1 //===- Writer.cpp ---------------------------------------------------------===//
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 #include "AArch64ErrataFix.h"
12 #include "CallGraphSort.h"
14 #include "Filesystem.h"
15 #include "LinkerScript.h"
17 #include "OutputSections.h"
18 #include "Relocations.h"
19 #include "SymbolTable.h"
21 #include "SyntheticSections.h"
23 #include "lld/Common/Memory.h"
24 #include "lld/Common/Strings.h"
25 #include "lld/Common/Threads.h"
26 #include "llvm/ADT/StringMap.h"
27 #include "llvm/ADT/StringSwitch.h"
31 using namespace llvm::ELF;
32 using namespace llvm::object;
33 using namespace llvm::support;
34 using namespace llvm::support::endian;
37 using namespace lld::elf;
40 // The writer writes a SymbolTable result to a file.
41 template <class ELFT> class Writer {
43 Writer() : Buffer(errorHandler().OutputBuffer) {}
44 typedef typename ELFT::Shdr Elf_Shdr;
45 typedef typename ELFT::Ehdr Elf_Ehdr;
46 typedef typename ELFT::Phdr Elf_Phdr;
51 void copyLocalSymbols();
52 void addSectionSymbols();
53 void forEachRelSec(llvm::function_ref<void(InputSectionBase &)> Fn);
55 void resolveShfLinkOrder();
56 void maybeAddThunks();
57 void sortInputSections();
58 void finalizeSections();
59 void checkExecuteOnly();
60 void setReservedSymbolSections();
62 std::vector<PhdrEntry *> createPhdrs();
63 void removeEmptyPTLoad();
64 void addPtArmExid(std::vector<PhdrEntry *> &Phdrs);
65 void assignFileOffsets();
66 void assignFileOffsetsBinary();
69 void fixSectionAlignments();
71 void writeTrapInstr();
74 void writeSectionsBinary();
77 std::unique_ptr<FileOutputBuffer> &Buffer;
79 void addRelIpltSymbols();
80 void addStartEndSymbols();
81 void addStartStopSymbols(OutputSection *Sec);
83 std::vector<PhdrEntry *> Phdrs;
86 uint64_t SectionHeaderOff;
88 } // anonymous namespace
90 static bool isSectionPrefix(StringRef Prefix, StringRef Name) {
91 return Name.startswith(Prefix) || Name == Prefix.drop_back();
94 StringRef elf::getOutputSectionName(const InputSectionBase *S) {
95 if (Config->Relocatable)
98 // This is for --emit-relocs. If .text.foo is emitted as .text.bar, we want
99 // to emit .rela.text.foo as .rela.text.bar for consistency (this is not
100 // technically required, but not doing it is odd). This code guarantees that.
101 if (auto *IS = dyn_cast<InputSection>(S)) {
102 if (InputSectionBase *Rel = IS->getRelocatedSection()) {
103 OutputSection *Out = Rel->getOutputSection();
104 if (S->Type == SHT_RELA)
105 return Saver.save(".rela" + Out->Name);
106 return Saver.save(".rel" + Out->Name);
110 // This check is for -z keep-text-section-prefix. This option separates text
111 // sections with prefix ".text.hot", ".text.unlikely", ".text.startup" or
113 // When enabled, this allows identifying the hot code region (.text.hot) in
114 // the final binary which can be selectively mapped to huge pages or mlocked,
116 if (Config->ZKeepTextSectionPrefix)
118 {".text.hot.", ".text.unlikely.", ".text.startup.", ".text.exit."})
119 if (isSectionPrefix(V, S->Name))
120 return V.drop_back();
123 {".text.", ".rodata.", ".data.rel.ro.", ".data.", ".bss.rel.ro.",
124 ".bss.", ".init_array.", ".fini_array.", ".ctors.", ".dtors.", ".tbss.",
125 ".gcc_except_table.", ".tdata.", ".ARM.exidx.", ".ARM.extab."})
126 if (isSectionPrefix(V, S->Name))
127 return V.drop_back();
129 // CommonSection is identified as "COMMON" in linker scripts.
130 // By default, it should go to .bss section.
131 if (S->Name == "COMMON")
137 static bool needsInterpSection() {
138 return !SharedFiles.empty() && !Config->DynamicLinker.empty() &&
139 Script->needsInterpSection();
142 template <class ELFT> void elf::writeResult() { Writer<ELFT>().run(); }
144 template <class ELFT> void Writer<ELFT>::removeEmptyPTLoad() {
145 llvm::erase_if(Phdrs, [&](const PhdrEntry *P) {
146 if (P->p_type != PT_LOAD)
150 uint64_t Size = P->LastSec->Addr + P->LastSec->Size - P->FirstSec->Addr;
155 template <class ELFT> static void combineEhFrameSections() {
156 for (InputSectionBase *&S : InputSections) {
157 EhInputSection *ES = dyn_cast<EhInputSection>(S);
158 if (!ES || !ES->Live)
161 In.EhFrame->addSection<ELFT>(ES);
165 std::vector<InputSectionBase *> &V = InputSections;
166 V.erase(std::remove(V.begin(), V.end(), nullptr), V.end());
169 static Defined *addOptionalRegular(StringRef Name, SectionBase *Sec,
170 uint64_t Val, uint8_t StOther = STV_HIDDEN,
171 uint8_t Binding = STB_GLOBAL) {
172 Symbol *S = Symtab->find(Name);
173 if (!S || S->isDefined())
175 return Symtab->addDefined(Name, StOther, STT_NOTYPE, Val,
176 /*Size=*/0, Binding, Sec,
180 static Defined *addAbsolute(StringRef Name) {
181 return Symtab->addDefined(Name, STV_HIDDEN, STT_NOTYPE, 0, 0, STB_GLOBAL,
185 // The linker is expected to define some symbols depending on
186 // the linking result. This function defines such symbols.
187 void elf::addReservedSymbols() {
188 if (Config->EMachine == EM_MIPS) {
189 // Define _gp for MIPS. st_value of _gp symbol will be updated by Writer
190 // so that it points to an absolute address which by default is relative
191 // to GOT. Default offset is 0x7ff0.
192 // See "Global Data Symbols" in Chapter 6 in the following document:
193 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
194 ElfSym::MipsGp = addAbsolute("_gp");
196 // On MIPS O32 ABI, _gp_disp is a magic symbol designates offset between
197 // start of function and 'gp' pointer into GOT.
198 if (Symtab->find("_gp_disp"))
199 ElfSym::MipsGpDisp = addAbsolute("_gp_disp");
201 // The __gnu_local_gp is a magic symbol equal to the current value of 'gp'
202 // pointer. This symbol is used in the code generated by .cpload pseudo-op
203 // in case of using -mno-shared option.
204 // https://sourceware.org/ml/binutils/2004-12/msg00094.html
205 if (Symtab->find("__gnu_local_gp"))
206 ElfSym::MipsLocalGp = addAbsolute("__gnu_local_gp");
209 // The Power Architecture 64-bit v2 ABI defines a TableOfContents (TOC) which
210 // combines the typical ELF GOT with the small data sections. It commonly
211 // includes .got .toc .sdata .sbss. The .TOC. symbol replaces both
212 // _GLOBAL_OFFSET_TABLE_ and _SDA_BASE_ from the 32-bit ABI. It is used to
213 // represent the TOC base which is offset by 0x8000 bytes from the start of
215 // We do not allow _GLOBAL_OFFSET_TABLE_ to be defined by input objects as the
216 // correctness of some relocations depends on its value.
217 StringRef GotTableSymName =
218 (Config->EMachine == EM_PPC64) ? ".TOC." : "_GLOBAL_OFFSET_TABLE_";
219 if (Symbol *S = Symtab->find(GotTableSymName)) {
221 error(toString(S->File) + " cannot redefine linker defined symbol '" +
222 GotTableSymName + "'");
224 ElfSym::GlobalOffsetTable = Symtab->addDefined(
225 GotTableSymName, STV_HIDDEN, STT_NOTYPE, Target->GotBaseSymOff,
226 /*Size=*/0, STB_GLOBAL, Out::ElfHeader,
230 // __ehdr_start is the location of ELF file headers. Note that we define
231 // this symbol unconditionally even when using a linker script, which
232 // differs from the behavior implemented by GNU linker which only define
233 // this symbol if ELF headers are in the memory mapped segment.
234 addOptionalRegular("__ehdr_start", Out::ElfHeader, 0, STV_HIDDEN);
236 // __executable_start is not documented, but the expectation of at
237 // least the Android libc is that it points to the ELF header.
238 addOptionalRegular("__executable_start", Out::ElfHeader, 0, STV_HIDDEN);
240 // __dso_handle symbol is passed to cxa_finalize as a marker to identify
241 // each DSO. The address of the symbol doesn't matter as long as they are
242 // different in different DSOs, so we chose the start address of the DSO.
243 addOptionalRegular("__dso_handle", Out::ElfHeader, 0, STV_HIDDEN);
245 // If linker script do layout we do not need to create any standart symbols.
246 if (Script->HasSectionsCommand)
249 auto Add = [](StringRef S, int64_t Pos) {
250 return addOptionalRegular(S, Out::ElfHeader, Pos, STV_DEFAULT);
253 ElfSym::Bss = Add("__bss_start", 0);
254 ElfSym::End1 = Add("end", -1);
255 ElfSym::End2 = Add("_end", -1);
256 ElfSym::Etext1 = Add("etext", -1);
257 ElfSym::Etext2 = Add("_etext", -1);
258 ElfSym::Edata1 = Add("edata", -1);
259 ElfSym::Edata2 = Add("_edata", -1);
262 static OutputSection *findSection(StringRef Name) {
263 for (BaseCommand *Base : Script->SectionCommands)
264 if (auto *Sec = dyn_cast<OutputSection>(Base))
265 if (Sec->Name == Name)
270 // Initialize Out members.
271 template <class ELFT> static void createSyntheticSections() {
272 // Initialize all pointers with NULL. This is needed because
273 // you can call lld::elf::main more than once as a library.
274 memset(&Out::First, 0, sizeof(Out));
276 auto Add = [](InputSectionBase *Sec) { InputSections.push_back(Sec); };
278 In.DynStrTab = make<StringTableSection>(".dynstr", true);
279 In.Dynamic = make<DynamicSection<ELFT>>();
280 if (Config->AndroidPackDynRelocs) {
281 In.RelaDyn = make<AndroidPackedRelocationSection<ELFT>>(
282 Config->IsRela ? ".rela.dyn" : ".rel.dyn");
284 In.RelaDyn = make<RelocationSection<ELFT>>(
285 Config->IsRela ? ".rela.dyn" : ".rel.dyn", Config->ZCombreloc);
287 In.ShStrTab = make<StringTableSection>(".shstrtab", false);
289 Out::ProgramHeaders = make<OutputSection>("", 0, SHF_ALLOC);
290 Out::ProgramHeaders->Alignment = Config->Wordsize;
292 if (needsInterpSection()) {
293 In.Interp = createInterpSection();
297 if (Config->Strip != StripPolicy::All) {
298 In.StrTab = make<StringTableSection>(".strtab", false);
299 In.SymTab = make<SymbolTableSection<ELFT>>(*In.StrTab);
300 In.SymTabShndx = make<SymtabShndxSection>();
303 if (Config->BuildId != BuildIdKind::None) {
304 In.BuildId = make<BuildIdSection>();
308 In.Bss = make<BssSection>(".bss", 0, 1);
311 // If there is a SECTIONS command and a .data.rel.ro section name use name
312 // .data.rel.ro.bss so that we match in the .data.rel.ro output section.
313 // This makes sure our relro is contiguous.
314 bool HasDataRelRo = Script->HasSectionsCommand && findSection(".data.rel.ro");
316 make<BssSection>(HasDataRelRo ? ".data.rel.ro.bss" : ".bss.rel.ro", 0, 1);
319 // Add MIPS-specific sections.
320 if (Config->EMachine == EM_MIPS) {
321 if (!Config->Shared && Config->HasDynSymTab) {
322 In.MipsRldMap = make<MipsRldMapSection>();
325 if (auto *Sec = MipsAbiFlagsSection<ELFT>::create())
327 if (auto *Sec = MipsOptionsSection<ELFT>::create())
329 if (auto *Sec = MipsReginfoSection<ELFT>::create())
333 if (Config->HasDynSymTab) {
334 In.DynSymTab = make<SymbolTableSection<ELFT>>(*In.DynStrTab);
337 InX<ELFT>::VerSym = make<VersionTableSection<ELFT>>();
338 Add(InX<ELFT>::VerSym);
340 if (!Config->VersionDefinitions.empty()) {
341 In.VerDef = make<VersionDefinitionSection>();
345 InX<ELFT>::VerNeed = make<VersionNeedSection<ELFT>>();
346 Add(InX<ELFT>::VerNeed);
348 if (Config->GnuHash) {
349 In.GnuHashTab = make<GnuHashTableSection>();
353 if (Config->SysvHash) {
354 In.HashTab = make<HashTableSection>();
363 if (Config->RelrPackDynRelocs) {
364 In.RelrDyn = make<RelrSection<ELFT>>();
368 // Add .got. MIPS' .got is so different from the other archs,
369 // it has its own class.
370 if (Config->EMachine == EM_MIPS) {
371 In.MipsGot = make<MipsGotSection>();
374 In.Got = make<GotSection>();
378 if (Config->EMachine == EM_PPC64) {
379 In.PPC64LongBranchTarget = make<PPC64LongBranchTargetSection>();
380 Add(In.PPC64LongBranchTarget);
383 In.GotPlt = make<GotPltSection>();
385 In.IgotPlt = make<IgotPltSection>();
388 if (Config->GdbIndex) {
389 In.GdbIndex = GdbIndexSection::create<ELFT>();
393 // We always need to add rel[a].plt to output if it has entries.
394 // Even for static linking it can contain R_[*]_IRELATIVE relocations.
395 In.RelaPlt = make<RelocationSection<ELFT>>(
396 Config->IsRela ? ".rela.plt" : ".rel.plt", false /*Sort*/);
399 // The RelaIplt immediately follows .rel.plt (.rel.dyn for ARM) to ensure
400 // that the IRelative relocations are processed last by the dynamic loader.
401 // We cannot place the iplt section in .rel.dyn when Android relocation
402 // packing is enabled because that would cause a section type mismatch.
403 // However, because the Android dynamic loader reads .rel.plt after .rel.dyn,
404 // we can get the desired behaviour by placing the iplt section in .rel.plt.
405 In.RelaIplt = make<RelocationSection<ELFT>>(
406 (Config->EMachine == EM_ARM && !Config->AndroidPackDynRelocs)
412 In.Plt = make<PltSection>(false);
414 In.Iplt = make<PltSection>(true);
417 // .note.GNU-stack is always added when we are creating a re-linkable
418 // object file. Other linkers are using the presence of this marker
419 // section to control the executable-ness of the stack area, but that
420 // is irrelevant these days. Stack area should always be non-executable
421 // by default. So we emit this section unconditionally.
422 if (Config->Relocatable)
423 Add(make<GnuStackSection>());
425 if (!Config->Relocatable) {
426 if (Config->EhFrameHdr) {
427 In.EhFrameHdr = make<EhFrameHeader>();
430 In.EhFrame = make<EhFrameSection>();
442 if (Config->EMachine == EM_ARM && !Config->Relocatable)
443 // Add a sentinel to terminate .ARM.exidx. It helps an unwinder
444 // to find the exact address range of the last entry.
445 Add(make<ARMExidxSentinelSection>());
448 // The main function of the writer.
449 template <class ELFT> void Writer<ELFT>::run() {
450 // Create linker-synthesized sections such as .got or .plt.
451 // Such sections are of type input section.
452 createSyntheticSections<ELFT>();
454 if (!Config->Relocatable)
455 combineEhFrameSections<ELFT>();
457 // We want to process linker script commands. When SECTIONS command
458 // is given we let it create sections.
459 Script->processSectionCommands();
461 // Linker scripts controls how input sections are assigned to output sections.
462 // Input sections that were not handled by scripts are called "orphans", and
463 // they are assigned to output sections by the default rule. Process that.
464 Script->addOrphanSections();
466 if (Config->Discard != DiscardPolicy::All)
469 if (Config->CopyRelocs)
472 // Now that we have a complete set of output sections. This function
473 // completes section contents. For example, we need to add strings
474 // to the string table, and add entries to .got and .plt.
475 // finalizeSections does that.
481 Script->assignAddresses();
483 // If -compressed-debug-sections is specified, we need to compress
484 // .debug_* sections. Do it right now because it changes the size of
486 for (OutputSection *Sec : OutputSections)
487 Sec->maybeCompress<ELFT>();
489 Script->allocateHeaders(Phdrs);
491 // Remove empty PT_LOAD to avoid causing the dynamic linker to try to mmap a
492 // 0 sized region. This has to be done late since only after assignAddresses
493 // we know the size of the sections.
496 if (!Config->OFormatBinary)
499 assignFileOffsetsBinary();
503 if (Config->Relocatable)
504 for (OutputSection *Sec : OutputSections)
507 if (Config->CheckSections)
510 // It does not make sense try to open the file if we have error already.
513 // Write the result down to a file.
518 if (!Config->OFormatBinary) {
523 writeSectionsBinary();
526 // Backfill .note.gnu.build-id section content. This is done at last
527 // because the content is usually a hash value of the entire output file.
532 // Handle -Map and -cref options.
534 writeCrossReferenceTable();
538 if (auto E = Buffer->commit())
539 error("failed to write to the output file: " + toString(std::move(E)));
542 static bool shouldKeepInSymtab(SectionBase *Sec, StringRef SymName,
547 if (Config->Discard == DiscardPolicy::None)
550 // In ELF assembly .L symbols are normally discarded by the assembler.
551 // If the assembler fails to do so, the linker discards them if
552 // * --discard-locals is used.
553 // * The symbol is in a SHF_MERGE section, which is normally the reason for
554 // the assembler keeping the .L symbol.
555 if (!SymName.startswith(".L") && !SymName.empty())
558 if (Config->Discard == DiscardPolicy::Locals)
561 return !Sec || !(Sec->Flags & SHF_MERGE);
564 static bool includeInSymtab(const Symbol &B) {
565 if (!B.isLocal() && !B.IsUsedInRegularObj)
568 if (auto *D = dyn_cast<Defined>(&B)) {
569 // Always include absolute symbols.
570 SectionBase *Sec = D->Section;
575 // Exclude symbols pointing to garbage-collected sections.
576 if (isa<InputSectionBase>(Sec) && !Sec->Live)
579 if (auto *S = dyn_cast<MergeInputSection>(Sec))
580 if (!S->getSectionPiece(D->Value)->Live)
587 // Local symbols are not in the linker's symbol table. This function scans
588 // each object file's symbol table to copy local symbols to the output.
589 template <class ELFT> void Writer<ELFT>::copyLocalSymbols() {
592 for (InputFile *File : ObjectFiles) {
593 ObjFile<ELFT> *F = cast<ObjFile<ELFT>>(File);
594 for (Symbol *B : F->getLocalSymbols()) {
597 ": broken object: getLocalSymbols returns a non-local symbol");
598 auto *DR = dyn_cast<Defined>(B);
600 // No reason to keep local undefined symbol in symtab.
603 if (!includeInSymtab(*B))
606 SectionBase *Sec = DR->Section;
607 if (!shouldKeepInSymtab(Sec, B->getName(), *B))
609 In.SymTab->addSymbol(B);
614 // Create a section symbol for each output section so that we can represent
615 // relocations that point to the section. If we know that no relocation is
616 // referring to a section (that happens if the section is a synthetic one), we
617 // don't create a section symbol for that section.
618 template <class ELFT> void Writer<ELFT>::addSectionSymbols() {
619 for (BaseCommand *Base : Script->SectionCommands) {
620 auto *Sec = dyn_cast<OutputSection>(Base);
623 auto I = llvm::find_if(Sec->SectionCommands, [](BaseCommand *Base) {
624 if (auto *ISD = dyn_cast<InputSectionDescription>(Base))
625 return !ISD->Sections.empty();
628 if (I == Sec->SectionCommands.end())
630 InputSection *IS = cast<InputSectionDescription>(*I)->Sections[0];
632 // Relocations are not using REL[A] section symbols.
633 if (IS->Type == SHT_REL || IS->Type == SHT_RELA)
636 // Unlike other synthetic sections, mergeable output sections contain data
637 // copied from input sections, and there may be a relocation pointing to its
638 // contents if -r or -emit-reloc are given.
639 if (isa<SyntheticSection>(IS) && !(IS->Flags & SHF_MERGE))
643 make<Defined>(IS->File, "", STB_LOCAL, /*StOther=*/0, STT_SECTION,
644 /*Value=*/0, /*Size=*/0, IS);
645 In.SymTab->addSymbol(Sym);
649 // Today's loaders have a feature to make segments read-only after
650 // processing dynamic relocations to enhance security. PT_GNU_RELRO
651 // is defined for that.
653 // This function returns true if a section needs to be put into a
654 // PT_GNU_RELRO segment.
655 static bool isRelroSection(const OutputSection *Sec) {
659 uint64_t Flags = Sec->Flags;
661 // Non-allocatable or non-writable sections don't need RELRO because
662 // they are not writable or not even mapped to memory in the first place.
663 // RELRO is for sections that are essentially read-only but need to
664 // be writable only at process startup to allow dynamic linker to
665 // apply relocations.
666 if (!(Flags & SHF_ALLOC) || !(Flags & SHF_WRITE))
669 // Once initialized, TLS data segments are used as data templates
670 // for a thread-local storage. For each new thread, runtime
671 // allocates memory for a TLS and copy templates there. No thread
672 // are supposed to use templates directly. Thus, it can be in RELRO.
676 // .init_array, .preinit_array and .fini_array contain pointers to
677 // functions that are executed on process startup or exit. These
678 // pointers are set by the static linker, and they are not expected
679 // to change at runtime. But if you are an attacker, you could do
680 // interesting things by manipulating pointers in .fini_array, for
681 // example. So they are put into RELRO.
682 uint32_t Type = Sec->Type;
683 if (Type == SHT_INIT_ARRAY || Type == SHT_FINI_ARRAY ||
684 Type == SHT_PREINIT_ARRAY)
687 // .got contains pointers to external symbols. They are resolved by
688 // the dynamic linker when a module is loaded into memory, and after
689 // that they are not expected to change. So, it can be in RELRO.
690 if (In.Got && Sec == In.Got->getParent())
693 // .toc is a GOT-ish section for PowerPC64. Their contents are accessed
694 // through r2 register, which is reserved for that purpose. Since r2 is used
695 // for accessing .got as well, .got and .toc need to be close enough in the
696 // virtual address space. Usually, .toc comes just after .got. Since we place
697 // .got into RELRO, .toc needs to be placed into RELRO too.
698 if (Sec->Name.equals(".toc"))
701 // .got.plt contains pointers to external function symbols. They are
702 // by default resolved lazily, so we usually cannot put it into RELRO.
703 // However, if "-z now" is given, the lazy symbol resolution is
704 // disabled, which enables us to put it into RELRO.
705 if (Sec == In.GotPlt->getParent())
708 // .dynamic section contains data for the dynamic linker, and
709 // there's no need to write to it at runtime, so it's better to put
711 if (Sec == In.Dynamic->getParent())
714 // Sections with some special names are put into RELRO. This is a
715 // bit unfortunate because section names shouldn't be significant in
716 // ELF in spirit. But in reality many linker features depend on
717 // magic section names.
718 StringRef S = Sec->Name;
719 return S == ".data.rel.ro" || S == ".bss.rel.ro" || S == ".ctors" ||
720 S == ".dtors" || S == ".jcr" || S == ".eh_frame" ||
721 S == ".openbsd.randomdata";
724 // We compute a rank for each section. The rank indicates where the
725 // section should be placed in the file. Instead of using simple
726 // numbers (0,1,2...), we use a series of flags. One for each decision
727 // point when placing the section.
728 // Using flags has two key properties:
729 // * It is easy to check if a give branch was taken.
730 // * It is easy two see how similar two ranks are (see getRankProximity).
732 RF_NOT_ADDR_SET = 1 << 18,
733 RF_NOT_ALLOC = 1 << 17,
734 RF_NOT_INTERP = 1 << 16,
735 RF_NOT_NOTE = 1 << 15,
737 RF_EXEC_WRITE = 1 << 13,
740 RF_NON_TLS_BSS = 1 << 10,
741 RF_NON_TLS_BSS_RO = 1 << 9,
744 RF_PPC_NOT_TOCBSS = 1 << 6,
745 RF_PPC_TOCL = 1 << 5,
748 RF_PPC_BRANCH_LT = 1 << 2,
749 RF_MIPS_GPREL = 1 << 1,
750 RF_MIPS_NOT_GOT = 1 << 0
753 static unsigned getSectionRank(const OutputSection *Sec) {
756 // We want to put section specified by -T option first, so we
757 // can start assigning VA starting from them later.
758 if (Config->SectionStartMap.count(Sec->Name))
760 Rank |= RF_NOT_ADDR_SET;
762 // Allocatable sections go first to reduce the total PT_LOAD size and
763 // so debug info doesn't change addresses in actual code.
764 if (!(Sec->Flags & SHF_ALLOC))
765 return Rank | RF_NOT_ALLOC;
767 // Put .interp first because some loaders want to see that section
768 // on the first page of the executable file when loaded into memory.
769 if (Sec->Name == ".interp")
771 Rank |= RF_NOT_INTERP;
773 // Put .note sections (which make up one PT_NOTE) at the beginning so that
774 // they are likely to be included in a core file even if core file size is
775 // limited. In particular, we want a .note.gnu.build-id and a .note.tag to be
776 // included in a core to match core files with executables.
777 if (Sec->Type == SHT_NOTE)
781 // Sort sections based on their access permission in the following
782 // order: R, RX, RWX, RW. This order is based on the following
784 // * Read-only sections come first such that they go in the
785 // PT_LOAD covering the program headers at the start of the file.
786 // * Read-only, executable sections come next.
787 // * Writable, executable sections follow such that .plt on
788 // architectures where it needs to be writable will be placed
789 // between .text and .data.
790 // * Writable sections come last, such that .bss lands at the very
791 // end of the last PT_LOAD.
792 bool IsExec = Sec->Flags & SHF_EXECINSTR;
793 bool IsWrite = Sec->Flags & SHF_WRITE;
797 Rank |= RF_EXEC_WRITE;
800 } else if (IsWrite) {
802 } else if (Sec->Type == SHT_PROGBITS) {
803 // Make non-executable and non-writable PROGBITS sections (e.g .rodata
804 // .eh_frame) closer to .text. They likely contain PC or GOT relative
805 // relocations and there could be relocation overflow if other huge sections
806 // (.dynstr .dynsym) were placed in between.
810 // If we got here we know that both A and B are in the same PT_LOAD.
812 bool IsTls = Sec->Flags & SHF_TLS;
813 bool IsNoBits = Sec->Type == SHT_NOBITS;
815 // The first requirement we have is to put (non-TLS) nobits sections last. The
816 // reason is that the only thing the dynamic linker will see about them is a
817 // p_memsz that is larger than p_filesz. Seeing that it zeros the end of the
818 // PT_LOAD, so that has to correspond to the nobits sections.
819 bool IsNonTlsNoBits = IsNoBits && !IsTls;
821 Rank |= RF_NON_TLS_BSS;
823 // We place nobits RelRo sections before plain r/w ones, and non-nobits RelRo
824 // sections after r/w ones, so that the RelRo sections are contiguous.
825 bool IsRelRo = isRelroSection(Sec);
826 if (IsNonTlsNoBits && !IsRelRo)
827 Rank |= RF_NON_TLS_BSS_RO;
828 if (!IsNonTlsNoBits && IsRelRo)
829 Rank |= RF_NON_TLS_BSS_RO;
831 // The TLS initialization block needs to be a single contiguous block in a R/W
832 // PT_LOAD, so stick TLS sections directly before the other RelRo R/W
833 // sections. The TLS NOBITS sections are placed here as they don't take up
834 // virtual address space in the PT_LOAD.
838 // Within the TLS initialization block, the non-nobits sections need to appear
843 // Some architectures have additional ordering restrictions for sections
844 // within the same PT_LOAD.
845 if (Config->EMachine == EM_PPC64) {
846 // PPC64 has a number of special SHT_PROGBITS+SHF_ALLOC+SHF_WRITE sections
847 // that we would like to make sure appear is a specific order to maximize
848 // their coverage by a single signed 16-bit offset from the TOC base
849 // pointer. Conversely, the special .tocbss section should be first among
850 // all SHT_NOBITS sections. This will put it next to the loaded special
851 // PPC64 sections (and, thus, within reach of the TOC base pointer).
852 StringRef Name = Sec->Name;
853 if (Name != ".tocbss")
854 Rank |= RF_PPC_NOT_TOCBSS;
865 if (Name == ".branch_lt")
866 Rank |= RF_PPC_BRANCH_LT;
869 if (Config->EMachine == EM_MIPS) {
870 // All sections with SHF_MIPS_GPREL flag should be grouped together
871 // because data in these sections is addressable with a gp relative address.
872 if (Sec->Flags & SHF_MIPS_GPREL)
873 Rank |= RF_MIPS_GPREL;
875 if (Sec->Name != ".got")
876 Rank |= RF_MIPS_NOT_GOT;
882 static bool compareSections(const BaseCommand *ACmd, const BaseCommand *BCmd) {
883 const OutputSection *A = cast<OutputSection>(ACmd);
884 const OutputSection *B = cast<OutputSection>(BCmd);
886 if (A->SortRank != B->SortRank)
887 return A->SortRank < B->SortRank;
889 if (!(A->SortRank & RF_NOT_ADDR_SET))
890 return Config->SectionStartMap.lookup(A->Name) <
891 Config->SectionStartMap.lookup(B->Name);
895 void PhdrEntry::add(OutputSection *Sec) {
899 p_align = std::max(p_align, Sec->Alignment);
900 if (p_type == PT_LOAD)
904 // The beginning and the ending of .rel[a].plt section are marked
905 // with __rel[a]_iplt_{start,end} symbols if it is a statically linked
906 // executable. The runtime needs these symbols in order to resolve
907 // all IRELATIVE relocs on startup. For dynamic executables, we don't
908 // need these symbols, since IRELATIVE relocs are resolved through GOT
909 // and PLT. For details, see http://www.airs.com/blog/archives/403.
910 template <class ELFT> void Writer<ELFT>::addRelIpltSymbols() {
911 if (Config->Relocatable || needsInterpSection())
914 // By default, __rela_iplt_{start,end} belong to a dummy section 0
915 // because .rela.plt might be empty and thus removed from output.
916 // We'll override Out::ElfHeader with In.RelaIplt later when we are
917 // sure that .rela.plt exists in output.
918 ElfSym::RelaIpltStart = addOptionalRegular(
919 Config->IsRela ? "__rela_iplt_start" : "__rel_iplt_start",
920 Out::ElfHeader, 0, STV_HIDDEN, STB_WEAK);
922 ElfSym::RelaIpltEnd = addOptionalRegular(
923 Config->IsRela ? "__rela_iplt_end" : "__rel_iplt_end",
924 Out::ElfHeader, 0, STV_HIDDEN, STB_WEAK);
927 template <class ELFT>
928 void Writer<ELFT>::forEachRelSec(
929 llvm::function_ref<void(InputSectionBase &)> Fn) {
930 // Scan all relocations. Each relocation goes through a series
931 // of tests to determine if it needs special treatment, such as
932 // creating GOT, PLT, copy relocations, etc.
933 // Note that relocations for non-alloc sections are directly
934 // processed by InputSection::relocateNonAlloc.
935 for (InputSectionBase *IS : InputSections)
936 if (IS->Live && isa<InputSection>(IS) && (IS->Flags & SHF_ALLOC))
938 for (EhInputSection *ES : In.EhFrame->Sections)
942 // This function generates assignments for predefined symbols (e.g. _end or
943 // _etext) and inserts them into the commands sequence to be processed at the
944 // appropriate time. This ensures that the value is going to be correct by the
945 // time any references to these symbols are processed and is equivalent to
946 // defining these symbols explicitly in the linker script.
947 template <class ELFT> void Writer<ELFT>::setReservedSymbolSections() {
948 if (ElfSym::GlobalOffsetTable) {
949 // The _GLOBAL_OFFSET_TABLE_ symbol is defined by target convention usually
950 // to the start of the .got or .got.plt section.
951 InputSection *GotSection = In.GotPlt;
952 if (!Target->GotBaseSymInGotPlt)
953 GotSection = In.MipsGot ? cast<InputSection>(In.MipsGot)
954 : cast<InputSection>(In.Got);
955 ElfSym::GlobalOffsetTable->Section = GotSection;
958 // .rela_iplt_{start,end} mark the start and the end of .rela.plt section.
959 if (ElfSym::RelaIpltStart && !In.RelaIplt->empty()) {
960 ElfSym::RelaIpltStart->Section = In.RelaIplt;
961 ElfSym::RelaIpltEnd->Section = In.RelaIplt;
962 ElfSym::RelaIpltEnd->Value = In.RelaIplt->getSize();
965 PhdrEntry *Last = nullptr;
966 PhdrEntry *LastRO = nullptr;
968 for (PhdrEntry *P : Phdrs) {
969 if (P->p_type != PT_LOAD)
972 if (!(P->p_flags & PF_W))
977 // _etext is the first location after the last read-only loadable segment.
979 ElfSym::Etext1->Section = LastRO->LastSec;
981 ElfSym::Etext2->Section = LastRO->LastSec;
985 // _edata points to the end of the last mapped initialized section.
986 OutputSection *Edata = nullptr;
987 for (OutputSection *OS : OutputSections) {
988 if (OS->Type != SHT_NOBITS)
990 if (OS == Last->LastSec)
995 ElfSym::Edata1->Section = Edata;
997 ElfSym::Edata2->Section = Edata;
999 // _end is the first location after the uninitialized data region.
1001 ElfSym::End1->Section = Last->LastSec;
1003 ElfSym::End2->Section = Last->LastSec;
1007 ElfSym::Bss->Section = findSection(".bss");
1009 // Setup MIPS _gp_disp/__gnu_local_gp symbols which should
1010 // be equal to the _gp symbol's value.
1011 if (ElfSym::MipsGp) {
1012 // Find GP-relative section with the lowest address
1013 // and use this address to calculate default _gp value.
1014 for (OutputSection *OS : OutputSections) {
1015 if (OS->Flags & SHF_MIPS_GPREL) {
1016 ElfSym::MipsGp->Section = OS;
1017 ElfSym::MipsGp->Value = 0x7ff0;
1024 // We want to find how similar two ranks are.
1025 // The more branches in getSectionRank that match, the more similar they are.
1026 // Since each branch corresponds to a bit flag, we can just use
1027 // countLeadingZeros.
1028 static int getRankProximityAux(OutputSection *A, OutputSection *B) {
1029 return countLeadingZeros(A->SortRank ^ B->SortRank);
1032 static int getRankProximity(OutputSection *A, BaseCommand *B) {
1033 if (auto *Sec = dyn_cast<OutputSection>(B))
1034 return getRankProximityAux(A, Sec);
1038 // When placing orphan sections, we want to place them after symbol assignments
1039 // so that an orphan after
1043 // doesn't break the intended meaning of the begin/end symbols.
1044 // We don't want to go over sections since findOrphanPos is the
1045 // one in charge of deciding the order of the sections.
1046 // We don't want to go over changes to '.', since doing so in
1047 // rx_sec : { *(rx_sec) }
1048 // . = ALIGN(0x1000);
1049 // /* The RW PT_LOAD starts here*/
1050 // rw_sec : { *(rw_sec) }
1051 // would mean that the RW PT_LOAD would become unaligned.
1052 static bool shouldSkip(BaseCommand *Cmd) {
1053 if (auto *Assign = dyn_cast<SymbolAssignment>(Cmd))
1054 return Assign->Name != ".";
1058 // We want to place orphan sections so that they share as much
1059 // characteristics with their neighbors as possible. For example, if
1060 // both are rw, or both are tls.
1061 template <typename ELFT>
1062 static std::vector<BaseCommand *>::iterator
1063 findOrphanPos(std::vector<BaseCommand *>::iterator B,
1064 std::vector<BaseCommand *>::iterator E) {
1065 OutputSection *Sec = cast<OutputSection>(*E);
1067 // Find the first element that has as close a rank as possible.
1068 auto I = std::max_element(B, E, [=](BaseCommand *A, BaseCommand *B) {
1069 return getRankProximity(Sec, A) < getRankProximity(Sec, B);
1074 // Consider all existing sections with the same proximity.
1075 int Proximity = getRankProximity(Sec, *I);
1076 for (; I != E; ++I) {
1077 auto *CurSec = dyn_cast<OutputSection>(*I);
1080 if (getRankProximity(Sec, CurSec) != Proximity ||
1081 Sec->SortRank < CurSec->SortRank)
1085 auto IsOutputSec = [](BaseCommand *Cmd) { return isa<OutputSection>(Cmd); };
1086 auto J = std::find_if(llvm::make_reverse_iterator(I),
1087 llvm::make_reverse_iterator(B), IsOutputSec);
1090 // As a special case, if the orphan section is the last section, put
1091 // it at the very end, past any other commands.
1092 // This matches bfd's behavior and is convenient when the linker script fully
1093 // specifies the start of the file, but doesn't care about the end (the non
1094 // alloc sections for example).
1095 auto NextSec = std::find_if(I, E, IsOutputSec);
1099 while (I != E && shouldSkip(*I))
1104 // Builds section order for handling --symbol-ordering-file.
1105 static DenseMap<const InputSectionBase *, int> buildSectionOrder() {
1106 DenseMap<const InputSectionBase *, int> SectionOrder;
1107 // Use the rarely used option -call-graph-ordering-file to sort sections.
1108 if (!Config->CallGraphProfile.empty())
1109 return computeCallGraphProfileOrder();
1111 if (Config->SymbolOrderingFile.empty())
1112 return SectionOrder;
1114 struct SymbolOrderEntry {
1119 // Build a map from symbols to their priorities. Symbols that didn't
1120 // appear in the symbol ordering file have the lowest priority 0.
1121 // All explicitly mentioned symbols have negative (higher) priorities.
1122 DenseMap<StringRef, SymbolOrderEntry> SymbolOrder;
1123 int Priority = -Config->SymbolOrderingFile.size();
1124 for (StringRef S : Config->SymbolOrderingFile)
1125 SymbolOrder.insert({S, {Priority++, false}});
1127 // Build a map from sections to their priorities.
1128 auto AddSym = [&](Symbol &Sym) {
1129 auto It = SymbolOrder.find(Sym.getName());
1130 if (It == SymbolOrder.end())
1132 SymbolOrderEntry &Ent = It->second;
1135 maybeWarnUnorderableSymbol(&Sym);
1137 if (auto *D = dyn_cast<Defined>(&Sym)) {
1138 if (auto *Sec = dyn_cast_or_null<InputSectionBase>(D->Section)) {
1139 int &Priority = SectionOrder[cast<InputSectionBase>(Sec->Repl)];
1140 Priority = std::min(Priority, Ent.Priority);
1145 // We want both global and local symbols. We get the global ones from the
1146 // symbol table and iterate the object files for the local ones.
1147 for (Symbol *Sym : Symtab->getSymbols())
1150 for (InputFile *File : ObjectFiles)
1151 for (Symbol *Sym : File->getSymbols())
1155 if (Config->WarnSymbolOrdering)
1156 for (auto OrderEntry : SymbolOrder)
1157 if (!OrderEntry.second.Present)
1158 warn("symbol ordering file: no such symbol: " + OrderEntry.first);
1160 return SectionOrder;
1163 // Sorts the sections in ISD according to the provided section order.
1165 sortISDBySectionOrder(InputSectionDescription *ISD,
1166 const DenseMap<const InputSectionBase *, int> &Order) {
1167 std::vector<InputSection *> UnorderedSections;
1168 std::vector<std::pair<InputSection *, int>> OrderedSections;
1169 uint64_t UnorderedSize = 0;
1171 for (InputSection *IS : ISD->Sections) {
1172 auto I = Order.find(IS);
1173 if (I == Order.end()) {
1174 UnorderedSections.push_back(IS);
1175 UnorderedSize += IS->getSize();
1178 OrderedSections.push_back({IS, I->second});
1180 llvm::sort(OrderedSections, [&](std::pair<InputSection *, int> A,
1181 std::pair<InputSection *, int> B) {
1182 return A.second < B.second;
1185 // Find an insertion point for the ordered section list in the unordered
1186 // section list. On targets with limited-range branches, this is the mid-point
1187 // of the unordered section list. This decreases the likelihood that a range
1188 // extension thunk will be needed to enter or exit the ordered region. If the
1189 // ordered section list is a list of hot functions, we can generally expect
1190 // the ordered functions to be called more often than the unordered functions,
1191 // making it more likely that any particular call will be within range, and
1192 // therefore reducing the number of thunks required.
1194 // For example, imagine that you have 8MB of hot code and 32MB of cold code.
1195 // If the layout is:
1200 // only the first 8-16MB of the cold code (depending on which hot function it
1201 // is actually calling) can call the hot code without a range extension thunk.
1202 // However, if we use this layout:
1208 // both the last 8-16MB of the first block of cold code and the first 8-16MB
1209 // of the second block of cold code can call the hot code without a thunk. So
1210 // we effectively double the amount of code that could potentially call into
1211 // the hot code without a thunk.
1213 if (Target->getThunkSectionSpacing() && !OrderedSections.empty()) {
1214 uint64_t UnorderedPos = 0;
1215 for (; InsPt != UnorderedSections.size(); ++InsPt) {
1216 UnorderedPos += UnorderedSections[InsPt]->getSize();
1217 if (UnorderedPos > UnorderedSize / 2)
1222 ISD->Sections.clear();
1223 for (InputSection *IS : makeArrayRef(UnorderedSections).slice(0, InsPt))
1224 ISD->Sections.push_back(IS);
1225 for (std::pair<InputSection *, int> P : OrderedSections)
1226 ISD->Sections.push_back(P.first);
1227 for (InputSection *IS : makeArrayRef(UnorderedSections).slice(InsPt))
1228 ISD->Sections.push_back(IS);
1231 static void sortSection(OutputSection *Sec,
1232 const DenseMap<const InputSectionBase *, int> &Order) {
1233 StringRef Name = Sec->Name;
1235 // Sort input sections by section name suffixes for
1236 // __attribute__((init_priority(N))).
1237 if (Name == ".init_array" || Name == ".fini_array") {
1238 if (!Script->HasSectionsCommand)
1239 Sec->sortInitFini();
1243 // Sort input sections by the special rule for .ctors and .dtors.
1244 if (Name == ".ctors" || Name == ".dtors") {
1245 if (!Script->HasSectionsCommand)
1246 Sec->sortCtorsDtors();
1250 // Never sort these.
1251 if (Name == ".init" || Name == ".fini")
1254 // Sort input sections by priority using the list provided
1255 // by --symbol-ordering-file.
1257 for (BaseCommand *B : Sec->SectionCommands)
1258 if (auto *ISD = dyn_cast<InputSectionDescription>(B))
1259 sortISDBySectionOrder(ISD, Order);
1262 // If no layout was provided by linker script, we want to apply default
1263 // sorting for special input sections. This also handles --symbol-ordering-file.
1264 template <class ELFT> void Writer<ELFT>::sortInputSections() {
1265 // Build the order once since it is expensive.
1266 DenseMap<const InputSectionBase *, int> Order = buildSectionOrder();
1267 for (BaseCommand *Base : Script->SectionCommands)
1268 if (auto *Sec = dyn_cast<OutputSection>(Base))
1269 sortSection(Sec, Order);
1272 template <class ELFT> void Writer<ELFT>::sortSections() {
1273 Script->adjustSectionsBeforeSorting();
1275 // Don't sort if using -r. It is not necessary and we want to preserve the
1276 // relative order for SHF_LINK_ORDER sections.
1277 if (Config->Relocatable)
1280 sortInputSections();
1282 for (BaseCommand *Base : Script->SectionCommands) {
1283 auto *OS = dyn_cast<OutputSection>(Base);
1286 OS->SortRank = getSectionRank(OS);
1288 // We want to assign rude approximation values to OutSecOff fields
1289 // to know the relative order of the input sections. We use it for
1290 // sorting SHF_LINK_ORDER sections. See resolveShfLinkOrder().
1292 for (InputSection *Sec : getInputSections(OS))
1293 Sec->OutSecOff = I++;
1296 if (!Script->HasSectionsCommand) {
1297 // We know that all the OutputSections are contiguous in this case.
1298 auto IsSection = [](BaseCommand *Base) { return isa<OutputSection>(Base); };
1300 llvm::find_if(Script->SectionCommands, IsSection),
1301 llvm::find_if(llvm::reverse(Script->SectionCommands), IsSection).base(),
1306 // Orphan sections are sections present in the input files which are
1307 // not explicitly placed into the output file by the linker script.
1309 // The sections in the linker script are already in the correct
1310 // order. We have to figuere out where to insert the orphan
1313 // The order of the sections in the script is arbitrary and may not agree with
1314 // compareSections. This means that we cannot easily define a strict weak
1315 // ordering. To see why, consider a comparison of a section in the script and
1316 // one not in the script. We have a two simple options:
1317 // * Make them equivalent (a is not less than b, and b is not less than a).
1318 // The problem is then that equivalence has to be transitive and we can
1319 // have sections a, b and c with only b in a script and a less than c
1320 // which breaks this property.
1321 // * Use compareSectionsNonScript. Given that the script order doesn't have
1322 // to match, we can end up with sections a, b, c, d where b and c are in the
1323 // script and c is compareSectionsNonScript less than b. In which case d
1324 // can be equivalent to c, a to b and d < a. As a concrete example:
1325 // .a (rx) # not in script
1326 // .b (rx) # in script
1327 // .c (ro) # in script
1328 // .d (ro) # not in script
1330 // The way we define an order then is:
1331 // * Sort only the orphan sections. They are in the end right now.
1332 // * Move each orphan section to its preferred position. We try
1333 // to put each section in the last position where it can share
1336 // There is some ambiguity as to where exactly a new entry should be
1337 // inserted, because Commands contains not only output section
1338 // commands but also other types of commands such as symbol assignment
1339 // expressions. There's no correct answer here due to the lack of the
1340 // formal specification of the linker script. We use heuristics to
1341 // determine whether a new output command should be added before or
1342 // after another commands. For the details, look at shouldSkip
1345 auto I = Script->SectionCommands.begin();
1346 auto E = Script->SectionCommands.end();
1347 auto NonScriptI = std::find_if(I, E, [](BaseCommand *Base) {
1348 if (auto *Sec = dyn_cast<OutputSection>(Base))
1349 return Sec->SectionIndex == UINT32_MAX;
1353 // Sort the orphan sections.
1354 std::stable_sort(NonScriptI, E, compareSections);
1356 // As a horrible special case, skip the first . assignment if it is before any
1357 // section. We do this because it is common to set a load address by starting
1358 // the script with ". = 0xabcd" and the expectation is that every section is
1360 auto FirstSectionOrDotAssignment =
1361 std::find_if(I, E, [](BaseCommand *Cmd) { return !shouldSkip(Cmd); });
1362 if (FirstSectionOrDotAssignment != E &&
1363 isa<SymbolAssignment>(**FirstSectionOrDotAssignment))
1364 ++FirstSectionOrDotAssignment;
1365 I = FirstSectionOrDotAssignment;
1367 while (NonScriptI != E) {
1368 auto Pos = findOrphanPos<ELFT>(I, NonScriptI);
1369 OutputSection *Orphan = cast<OutputSection>(*NonScriptI);
1371 // As an optimization, find all sections with the same sort rank
1372 // and insert them with one rotate.
1373 unsigned Rank = Orphan->SortRank;
1374 auto End = std::find_if(NonScriptI + 1, E, [=](BaseCommand *Cmd) {
1375 return cast<OutputSection>(Cmd)->SortRank != Rank;
1377 std::rotate(Pos, NonScriptI, End);
1381 Script->adjustSectionsAfterSorting();
1384 static bool compareByFilePosition(InputSection *A, InputSection *B) {
1385 // Synthetic, i. e. a sentinel section, should go last.
1386 if (A->kind() == InputSectionBase::Synthetic ||
1387 B->kind() == InputSectionBase::Synthetic)
1388 return A->kind() != InputSectionBase::Synthetic;
1390 InputSection *LA = A->getLinkOrderDep();
1391 InputSection *LB = B->getLinkOrderDep();
1392 OutputSection *AOut = LA->getParent();
1393 OutputSection *BOut = LB->getParent();
1396 return AOut->SectionIndex < BOut->SectionIndex;
1397 return LA->OutSecOff < LB->OutSecOff;
1400 // This function is used by the --merge-exidx-entries to detect duplicate
1401 // .ARM.exidx sections. It is Arm only.
1403 // The .ARM.exidx section is of the form:
1404 // | PREL31 offset to function | Unwind instructions for function |
1405 // where the unwind instructions are either a small number of unwind
1406 // instructions inlined into the table entry, the special CANT_UNWIND value of
1407 // 0x1 or a PREL31 offset into a .ARM.extab Section that contains unwind
1410 // We return true if all the unwind instructions in the .ARM.exidx entries of
1411 // Cur can be merged into the last entry of Prev.
1412 static bool isDuplicateArmExidxSec(InputSection *Prev, InputSection *Cur) {
1414 // References to .ARM.Extab Sections have bit 31 clear and are not the
1415 // special EXIDX_CANTUNWIND bit-pattern.
1416 auto IsExtabRef = [](uint32_t Unwind) {
1417 return (Unwind & 0x80000000) == 0 && Unwind != 0x1;
1425 // Get the last table Entry from the previous .ARM.exidx section.
1426 const ExidxEntry &PrevEntry = Prev->getDataAs<ExidxEntry>().back();
1427 if (IsExtabRef(PrevEntry.Unwind))
1430 // We consider the unwind instructions of an .ARM.exidx table entry
1431 // a duplicate if the previous unwind instructions if:
1432 // - Both are the special EXIDX_CANTUNWIND.
1433 // - Both are the same inline unwind instructions.
1434 // We do not attempt to follow and check links into .ARM.extab tables as
1435 // consecutive identical entries are rare and the effort to check that they
1436 // are identical is high.
1438 for (const ExidxEntry Entry : Cur->getDataAs<ExidxEntry>())
1439 if (IsExtabRef(Entry.Unwind) || Entry.Unwind != PrevEntry.Unwind)
1442 // All table entries in this .ARM.exidx Section can be merged into the
1443 // previous Section.
1447 template <class ELFT> void Writer<ELFT>::resolveShfLinkOrder() {
1448 for (OutputSection *Sec : OutputSections) {
1449 if (!(Sec->Flags & SHF_LINK_ORDER))
1452 // Link order may be distributed across several InputSectionDescriptions
1453 // but sort must consider them all at once.
1454 std::vector<InputSection **> ScriptSections;
1455 std::vector<InputSection *> Sections;
1456 for (BaseCommand *Base : Sec->SectionCommands) {
1457 if (auto *ISD = dyn_cast<InputSectionDescription>(Base)) {
1458 for (InputSection *&IS : ISD->Sections) {
1459 ScriptSections.push_back(&IS);
1460 Sections.push_back(IS);
1464 std::stable_sort(Sections.begin(), Sections.end(), compareByFilePosition);
1466 if (!Config->Relocatable && Config->EMachine == EM_ARM &&
1467 Sec->Type == SHT_ARM_EXIDX) {
1469 if (auto *Sentinel = dyn_cast<ARMExidxSentinelSection>(Sections.back())) {
1470 assert(Sections.size() >= 2 &&
1471 "We should create a sentinel section only if there are "
1472 "alive regular exidx sections.");
1474 // The last executable section is required to fill the sentinel.
1475 // Remember it here so that we don't have to find it again.
1476 Sentinel->Highest = Sections[Sections.size() - 2]->getLinkOrderDep();
1479 // The EHABI for the Arm Architecture permits consecutive identical
1480 // table entries to be merged. We use a simple implementation that
1481 // removes a .ARM.exidx Input Section if it can be merged into the
1482 // previous one. This does not require any rewriting of InputSection
1483 // contents but misses opportunities for fine grained deduplication
1484 // where only a subset of the InputSection contents can be merged.
1485 if (Config->MergeArmExidx) {
1487 // The last one is a sentinel entry which should not be removed.
1488 for (size_t I = 1; I < Sections.size() - 1; ++I) {
1489 if (isDuplicateArmExidxSec(Sections[Prev], Sections[I]))
1490 Sections[I] = nullptr;
1497 for (int I = 0, N = Sections.size(); I < N; ++I)
1498 *ScriptSections[I] = Sections[I];
1500 // Remove the Sections we marked as duplicate earlier.
1501 for (BaseCommand *Base : Sec->SectionCommands)
1502 if (auto *ISD = dyn_cast<InputSectionDescription>(Base))
1503 llvm::erase_if(ISD->Sections, [](InputSection *IS) { return !IS; });
1507 // For most RISC ISAs, we need to generate content that depends on the address
1508 // of InputSections. For example some architectures such as AArch64 use small
1509 // displacements for jump instructions that is the linker's responsibility for
1510 // creating range extension thunks for. As the generation of the content may
1511 // also alter InputSection addresses we must converge to a fixed point.
1512 template <class ELFT> void Writer<ELFT>::maybeAddThunks() {
1513 if (!Target->NeedsThunks && !Config->AndroidPackDynRelocs &&
1514 !Config->RelrPackDynRelocs)
1518 AArch64Err843419Patcher A64P;
1521 bool Changed = false;
1523 Script->assignAddresses();
1525 if (Target->NeedsThunks)
1526 Changed |= TC.createThunks(OutputSections);
1528 if (Config->FixCortexA53Errata843419) {
1530 Script->assignAddresses();
1531 Changed |= A64P.createFixes();
1535 In.MipsGot->updateAllocSize();
1537 Changed |= In.RelaDyn->updateAllocSize();
1540 Changed |= In.RelrDyn->updateAllocSize();
1547 static void finalizeSynthetic(SyntheticSection *Sec) {
1548 if (Sec && !Sec->empty() && Sec->getParent())
1549 Sec->finalizeContents();
1552 // In order to allow users to manipulate linker-synthesized sections,
1553 // we had to add synthetic sections to the input section list early,
1554 // even before we make decisions whether they are needed. This allows
1555 // users to write scripts like this: ".mygot : { .got }".
1557 // Doing it has an unintended side effects. If it turns out that we
1558 // don't need a .got (for example) at all because there's no
1559 // relocation that needs a .got, we don't want to emit .got.
1561 // To deal with the above problem, this function is called after
1562 // scanRelocations is called to remove synthetic sections that turn
1564 static void removeUnusedSyntheticSections() {
1565 // All input synthetic sections that can be empty are placed after
1566 // all regular ones. We iterate over them all and exit at first
1568 for (InputSectionBase *S : llvm::reverse(InputSections)) {
1569 SyntheticSection *SS = dyn_cast<SyntheticSection>(S);
1572 OutputSection *OS = SS->getParent();
1573 if (!OS || !SS->empty())
1576 // If we reach here, then SS is an unused synthetic section and we want to
1577 // remove it from corresponding input section description of output section.
1578 for (BaseCommand *B : OS->SectionCommands)
1579 if (auto *ISD = dyn_cast<InputSectionDescription>(B))
1580 llvm::erase_if(ISD->Sections,
1581 [=](InputSection *IS) { return IS == SS; });
1585 // Returns true if a symbol can be replaced at load-time by a symbol
1586 // with the same name defined in other ELF executable or DSO.
1587 static bool computeIsPreemptible(const Symbol &B) {
1588 assert(!B.isLocal());
1590 // Only symbols that appear in dynsym can be preempted.
1591 if (!B.includeInDynsym())
1594 // Only default visibility symbols can be preempted.
1595 if (B.Visibility != STV_DEFAULT)
1598 // At this point copy relocations have not been created yet, so any
1599 // symbol that is not defined locally is preemptible.
1603 // If we have a dynamic list it specifies which local symbols are preemptible.
1604 if (Config->HasDynamicList)
1607 if (!Config->Shared)
1610 // -Bsymbolic means that definitions are not preempted.
1611 if (Config->Bsymbolic || (Config->BsymbolicFunctions && B.isFunc()))
1616 // Create output section objects and add them to OutputSections.
1617 template <class ELFT> void Writer<ELFT>::finalizeSections() {
1618 Out::PreinitArray = findSection(".preinit_array");
1619 Out::InitArray = findSection(".init_array");
1620 Out::FiniArray = findSection(".fini_array");
1622 // The linker needs to define SECNAME_start, SECNAME_end and SECNAME_stop
1623 // symbols for sections, so that the runtime can get the start and end
1624 // addresses of each section by section name. Add such symbols.
1625 if (!Config->Relocatable) {
1626 addStartEndSymbols();
1627 for (BaseCommand *Base : Script->SectionCommands)
1628 if (auto *Sec = dyn_cast<OutputSection>(Base))
1629 addStartStopSymbols(Sec);
1632 // Add _DYNAMIC symbol. Unlike GNU gold, our _DYNAMIC symbol has no type.
1633 // It should be okay as no one seems to care about the type.
1634 // Even the author of gold doesn't remember why gold behaves that way.
1635 // https://sourceware.org/ml/binutils/2002-03/msg00360.html
1636 if (In.Dynamic->Parent)
1637 Symtab->addDefined("_DYNAMIC", STV_HIDDEN, STT_NOTYPE, 0 /*Value*/,
1638 /*Size=*/0, STB_WEAK, In.Dynamic,
1641 // Define __rel[a]_iplt_{start,end} symbols if needed.
1642 addRelIpltSymbols();
1644 // RISC-V's gp can address +/- 2 KiB, set it to .sdata + 0x800 if not defined.
1645 if (Config->EMachine == EM_RISCV)
1646 if (!dyn_cast_or_null<Defined>(Symtab->find("__global_pointer$")))
1647 addOptionalRegular("__global_pointer$", findSection(".sdata"), 0x800);
1649 // This responsible for splitting up .eh_frame section into
1650 // pieces. The relocation scan uses those pieces, so this has to be
1652 finalizeSynthetic(In.EhFrame);
1654 for (Symbol *S : Symtab->getSymbols()) {
1655 if (!S->IsPreemptible)
1656 S->IsPreemptible = computeIsPreemptible(*S);
1657 if (S->isGnuIFunc() && Config->ZIfuncnoplt)
1658 S->ExportDynamic = true;
1661 // Scan relocations. This must be done after every symbol is declared so that
1662 // we can correctly decide if a dynamic relocation is needed.
1663 if (!Config->Relocatable)
1664 forEachRelSec(scanRelocations<ELFT>);
1666 if (In.Plt && !In.Plt->empty())
1667 In.Plt->addSymbols();
1668 if (In.Iplt && !In.Iplt->empty())
1669 In.Iplt->addSymbols();
1671 // Now that we have defined all possible global symbols including linker-
1672 // synthesized ones. Visit all symbols to give the finishing touches.
1673 for (Symbol *Sym : Symtab->getSymbols()) {
1674 if (!includeInSymtab(*Sym))
1677 In.SymTab->addSymbol(Sym);
1679 if (Sym->includeInDynsym()) {
1680 In.DynSymTab->addSymbol(Sym);
1681 if (auto *File = dyn_cast_or_null<SharedFile<ELFT>>(Sym->File))
1682 if (File->IsNeeded && !Sym->isUndefined())
1683 InX<ELFT>::VerNeed->addSymbol(Sym);
1687 // Do not proceed if there was an undefined symbol.
1692 In.MipsGot->build<ELFT>();
1694 removeUnusedSyntheticSections();
1698 // Now that we have the final list, create a list of all the
1699 // OutputSections for convenience.
1700 for (BaseCommand *Base : Script->SectionCommands)
1701 if (auto *Sec = dyn_cast<OutputSection>(Base))
1702 OutputSections.push_back(Sec);
1704 // Prefer command line supplied address over other constraints.
1705 for (OutputSection *Sec : OutputSections) {
1706 auto I = Config->SectionStartMap.find(Sec->Name);
1707 if (I != Config->SectionStartMap.end())
1708 Sec->AddrExpr = [=] { return I->second; };
1711 // This is a bit of a hack. A value of 0 means undef, so we set it
1712 // to 1 to make __ehdr_start defined. The section number is not
1713 // particularly relevant.
1714 Out::ElfHeader->SectionIndex = 1;
1716 for (size_t I = 0, E = OutputSections.size(); I != E; ++I) {
1717 OutputSection *Sec = OutputSections[I];
1718 Sec->SectionIndex = I + 1;
1719 Sec->ShName = In.ShStrTab->addString(Sec->Name);
1722 // Binary and relocatable output does not have PHDRS.
1723 // The headers have to be created before finalize as that can influence the
1724 // image base and the dynamic section on mips includes the image base.
1725 if (!Config->Relocatable && !Config->OFormatBinary) {
1726 Phdrs = Script->hasPhdrsCommands() ? Script->createPhdrs() : createPhdrs();
1727 addPtArmExid(Phdrs);
1728 Out::ProgramHeaders->Size = sizeof(Elf_Phdr) * Phdrs.size();
1730 // Find the TLS segment. This happens before the section layout loop so that
1731 // Android relocation packing can look up TLS symbol addresses.
1732 for (PhdrEntry *P : Phdrs)
1733 if (P->p_type == PT_TLS)
1737 // Some symbols are defined in term of program headers. Now that we
1738 // have the headers, we can find out which sections they point to.
1739 setReservedSymbolSections();
1741 // Dynamic section must be the last one in this list and dynamic
1742 // symbol table section (DynSymTab) must be the first one.
1743 finalizeSynthetic(In.DynSymTab);
1744 finalizeSynthetic(In.Bss);
1745 finalizeSynthetic(In.BssRelRo);
1746 finalizeSynthetic(In.GnuHashTab);
1747 finalizeSynthetic(In.HashTab);
1748 finalizeSynthetic(In.SymTabShndx);
1749 finalizeSynthetic(In.ShStrTab);
1750 finalizeSynthetic(In.StrTab);
1751 finalizeSynthetic(In.VerDef);
1752 finalizeSynthetic(In.DynStrTab);
1753 finalizeSynthetic(In.Got);
1754 finalizeSynthetic(In.MipsGot);
1755 finalizeSynthetic(In.IgotPlt);
1756 finalizeSynthetic(In.GotPlt);
1757 finalizeSynthetic(In.RelaDyn);
1758 finalizeSynthetic(In.RelrDyn);
1759 finalizeSynthetic(In.RelaIplt);
1760 finalizeSynthetic(In.RelaPlt);
1761 finalizeSynthetic(In.Plt);
1762 finalizeSynthetic(In.Iplt);
1763 finalizeSynthetic(In.EhFrameHdr);
1764 finalizeSynthetic(InX<ELFT>::VerSym);
1765 finalizeSynthetic(InX<ELFT>::VerNeed);
1766 finalizeSynthetic(In.Dynamic);
1768 if (!Script->HasSectionsCommand && !Config->Relocatable)
1769 fixSectionAlignments();
1771 // After link order processing .ARM.exidx sections can be deduplicated, which
1772 // needs to be resolved before any other address dependent operation.
1773 resolveShfLinkOrder();
1775 // Jump instructions in many ISAs have small displacements, and therefore they
1776 // cannot jump to arbitrary addresses in memory. For example, RISC-V JAL
1777 // instruction can target only +-1 MiB from PC. It is a linker's
1778 // responsibility to create and insert small pieces of code between sections
1779 // to extend the ranges if jump targets are out of range. Such code pieces are
1782 // We add thunks at this stage. We couldn't do this before this point because
1783 // this is the earliest point where we know sizes of sections and their
1784 // layouts (that are needed to determine if jump targets are in range).
1787 // maybeAddThunks may have added local symbols to the static symbol table.
1788 finalizeSynthetic(In.SymTab);
1789 finalizeSynthetic(In.PPC64LongBranchTarget);
1791 // Fill other section headers. The dynamic table is finalized
1792 // at the end because some tags like RELSZ depend on result
1793 // of finalizing other sections.
1794 for (OutputSection *Sec : OutputSections)
1795 Sec->finalize<ELFT>();
1798 // Ensure data sections are not mixed with executable sections when
1799 // -execute-only is used. -execute-only is a feature to make pages executable
1800 // but not readable, and the feature is currently supported only on AArch64.
1801 template <class ELFT> void Writer<ELFT>::checkExecuteOnly() {
1802 if (!Config->ExecuteOnly)
1805 for (OutputSection *OS : OutputSections)
1806 if (OS->Flags & SHF_EXECINSTR)
1807 for (InputSection *IS : getInputSections(OS))
1808 if (!(IS->Flags & SHF_EXECINSTR))
1809 error("cannot place " + toString(IS) + " into " + toString(OS->Name) +
1810 ": -execute-only does not support intermingling data and code");
1813 // The linker is expected to define SECNAME_start and SECNAME_end
1814 // symbols for a few sections. This function defines them.
1815 template <class ELFT> void Writer<ELFT>::addStartEndSymbols() {
1816 // If a section does not exist, there's ambiguity as to how we
1817 // define _start and _end symbols for an init/fini section. Since
1818 // the loader assume that the symbols are always defined, we need to
1819 // always define them. But what value? The loader iterates over all
1820 // pointers between _start and _end to run global ctors/dtors, so if
1821 // the section is empty, their symbol values don't actually matter
1822 // as long as _start and _end point to the same location.
1824 // That said, we don't want to set the symbols to 0 (which is
1825 // probably the simplest value) because that could cause some
1826 // program to fail to link due to relocation overflow, if their
1827 // program text is above 2 GiB. We use the address of the .text
1828 // section instead to prevent that failure.
1830 // In a rare sitaution, .text section may not exist. If that's the
1831 // case, use the image base address as a last resort.
1832 OutputSection *Default = findSection(".text");
1834 Default = Out::ElfHeader;
1836 auto Define = [=](StringRef Start, StringRef End, OutputSection *OS) {
1838 addOptionalRegular(Start, OS, 0);
1839 addOptionalRegular(End, OS, -1);
1841 addOptionalRegular(Start, Default, 0);
1842 addOptionalRegular(End, Default, 0);
1846 Define("__preinit_array_start", "__preinit_array_end", Out::PreinitArray);
1847 Define("__init_array_start", "__init_array_end", Out::InitArray);
1848 Define("__fini_array_start", "__fini_array_end", Out::FiniArray);
1850 if (OutputSection *Sec = findSection(".ARM.exidx"))
1851 Define("__exidx_start", "__exidx_end", Sec);
1854 // If a section name is valid as a C identifier (which is rare because of
1855 // the leading '.'), linkers are expected to define __start_<secname> and
1856 // __stop_<secname> symbols. They are at beginning and end of the section,
1857 // respectively. This is not requested by the ELF standard, but GNU ld and
1858 // gold provide the feature, and used by many programs.
1859 template <class ELFT>
1860 void Writer<ELFT>::addStartStopSymbols(OutputSection *Sec) {
1861 StringRef S = Sec->Name;
1862 if (!isValidCIdentifier(S))
1864 addOptionalRegular(Saver.save("__start_" + S), Sec, 0, STV_PROTECTED);
1865 addOptionalRegular(Saver.save("__stop_" + S), Sec, -1, STV_PROTECTED);
1868 static bool needsPtLoad(OutputSection *Sec) {
1869 if (!(Sec->Flags & SHF_ALLOC) || Sec->Noload)
1872 // Don't allocate VA space for TLS NOBITS sections. The PT_TLS PHDR is
1873 // responsible for allocating space for them, not the PT_LOAD that
1874 // contains the TLS initialization image.
1875 if ((Sec->Flags & SHF_TLS) && Sec->Type == SHT_NOBITS)
1880 // Linker scripts are responsible for aligning addresses. Unfortunately, most
1881 // linker scripts are designed for creating two PT_LOADs only, one RX and one
1882 // RW. This means that there is no alignment in the RO to RX transition and we
1883 // cannot create a PT_LOAD there.
1884 static uint64_t computeFlags(uint64_t Flags) {
1886 return PF_R | PF_W | PF_X;
1887 if (Config->ExecuteOnly && (Flags & PF_X))
1888 return Flags & ~PF_R;
1889 if (Config->SingleRoRx && !(Flags & PF_W))
1890 return Flags | PF_X;
1894 // Decide which program headers to create and which sections to include in each
1896 template <class ELFT> std::vector<PhdrEntry *> Writer<ELFT>::createPhdrs() {
1897 std::vector<PhdrEntry *> Ret;
1898 auto AddHdr = [&](unsigned Type, unsigned Flags) -> PhdrEntry * {
1899 Ret.push_back(make<PhdrEntry>(Type, Flags));
1903 // The first phdr entry is PT_PHDR which describes the program header itself.
1904 AddHdr(PT_PHDR, PF_R)->add(Out::ProgramHeaders);
1906 // PT_INTERP must be the second entry if exists.
1907 if (OutputSection *Cmd = findSection(".interp"))
1908 AddHdr(PT_INTERP, Cmd->getPhdrFlags())->add(Cmd);
1910 // Add the first PT_LOAD segment for regular output sections.
1911 uint64_t Flags = computeFlags(PF_R);
1912 PhdrEntry *Load = AddHdr(PT_LOAD, Flags);
1914 // Add the headers. We will remove them if they don't fit.
1915 Load->add(Out::ElfHeader);
1916 Load->add(Out::ProgramHeaders);
1918 for (OutputSection *Sec : OutputSections) {
1919 if (!(Sec->Flags & SHF_ALLOC))
1921 if (!needsPtLoad(Sec))
1924 // Segments are contiguous memory regions that has the same attributes
1925 // (e.g. executable or writable). There is one phdr for each segment.
1926 // Therefore, we need to create a new phdr when the next section has
1927 // different flags or is loaded at a discontiguous address or memory
1928 // region using AT or AT> linker script command, respectively. At the same
1929 // time, we don't want to create a separate load segment for the headers,
1930 // even if the first output section has an AT or AT> attribute.
1931 uint64_t NewFlags = computeFlags(Sec->getPhdrFlags());
1932 if (((Sec->LMAExpr ||
1933 (Sec->LMARegion && (Sec->LMARegion != Load->FirstSec->LMARegion))) &&
1934 Load->LastSec != Out::ProgramHeaders) ||
1935 Sec->MemRegion != Load->FirstSec->MemRegion || Flags != NewFlags) {
1937 Load = AddHdr(PT_LOAD, NewFlags);
1944 // Add a TLS segment if any.
1945 PhdrEntry *TlsHdr = make<PhdrEntry>(PT_TLS, PF_R);
1946 for (OutputSection *Sec : OutputSections)
1947 if (Sec->Flags & SHF_TLS)
1949 if (TlsHdr->FirstSec)
1950 Ret.push_back(TlsHdr);
1952 // Add an entry for .dynamic.
1953 if (OutputSection *Sec = In.Dynamic->getParent())
1954 AddHdr(PT_DYNAMIC, Sec->getPhdrFlags())->add(Sec);
1956 // PT_GNU_RELRO includes all sections that should be marked as
1957 // read-only by dynamic linker after proccessing relocations.
1958 // Current dynamic loaders only support one PT_GNU_RELRO PHDR, give
1959 // an error message if more than one PT_GNU_RELRO PHDR is required.
1960 PhdrEntry *RelRo = make<PhdrEntry>(PT_GNU_RELRO, PF_R);
1961 bool InRelroPhdr = false;
1962 bool IsRelroFinished = false;
1963 for (OutputSection *Sec : OutputSections) {
1964 if (!needsPtLoad(Sec))
1966 if (isRelroSection(Sec)) {
1968 if (!IsRelroFinished)
1971 error("section: " + Sec->Name + " is not contiguous with other relro" +
1973 } else if (InRelroPhdr) {
1974 InRelroPhdr = false;
1975 IsRelroFinished = true;
1978 if (RelRo->FirstSec)
1979 Ret.push_back(RelRo);
1981 // PT_GNU_EH_FRAME is a special section pointing on .eh_frame_hdr.
1982 if (!In.EhFrame->empty() && In.EhFrameHdr && In.EhFrame->getParent() &&
1983 In.EhFrameHdr->getParent())
1984 AddHdr(PT_GNU_EH_FRAME, In.EhFrameHdr->getParent()->getPhdrFlags())
1985 ->add(In.EhFrameHdr->getParent());
1987 // PT_OPENBSD_RANDOMIZE is an OpenBSD-specific feature. That makes
1988 // the dynamic linker fill the segment with random data.
1989 if (OutputSection *Cmd = findSection(".openbsd.randomdata"))
1990 AddHdr(PT_OPENBSD_RANDOMIZE, Cmd->getPhdrFlags())->add(Cmd);
1992 // PT_GNU_STACK is a special section to tell the loader to make the
1993 // pages for the stack non-executable. If you really want an executable
1994 // stack, you can pass -z execstack, but that's not recommended for
1995 // security reasons.
1996 unsigned Perm = PF_R | PF_W;
1997 if (Config->ZExecstack)
1999 AddHdr(PT_GNU_STACK, Perm)->p_memsz = Config->ZStackSize;
2001 // PT_OPENBSD_WXNEEDED is a OpenBSD-specific header to mark the executable
2002 // is expected to perform W^X violations, such as calling mprotect(2) or
2003 // mmap(2) with PROT_WRITE | PROT_EXEC, which is prohibited by default on
2005 if (Config->ZWxneeded)
2006 AddHdr(PT_OPENBSD_WXNEEDED, PF_X);
2008 // Create one PT_NOTE per a group of contiguous .note sections.
2009 PhdrEntry *Note = nullptr;
2010 for (OutputSection *Sec : OutputSections) {
2011 if (Sec->Type == SHT_NOTE && (Sec->Flags & SHF_ALLOC)) {
2012 if (!Note || Sec->LMAExpr)
2013 Note = AddHdr(PT_NOTE, PF_R);
2022 template <class ELFT>
2023 void Writer<ELFT>::addPtArmExid(std::vector<PhdrEntry *> &Phdrs) {
2024 if (Config->EMachine != EM_ARM)
2026 auto I = llvm::find_if(OutputSections, [](OutputSection *Cmd) {
2027 return Cmd->Type == SHT_ARM_EXIDX;
2029 if (I == OutputSections.end())
2032 // PT_ARM_EXIDX is the ARM EHABI equivalent of PT_GNU_EH_FRAME
2033 PhdrEntry *ARMExidx = make<PhdrEntry>(PT_ARM_EXIDX, PF_R);
2035 Phdrs.push_back(ARMExidx);
2038 // The first section of each PT_LOAD, the first section in PT_GNU_RELRO and the
2039 // first section after PT_GNU_RELRO have to be page aligned so that the dynamic
2040 // linker can set the permissions.
2041 template <class ELFT> void Writer<ELFT>::fixSectionAlignments() {
2042 auto PageAlign = [](OutputSection *Cmd) {
2043 if (Cmd && !Cmd->AddrExpr)
2044 Cmd->AddrExpr = [=] {
2045 return alignTo(Script->getDot(), Config->MaxPageSize);
2049 for (const PhdrEntry *P : Phdrs)
2050 if (P->p_type == PT_LOAD && P->FirstSec)
2051 PageAlign(P->FirstSec);
2053 for (const PhdrEntry *P : Phdrs) {
2054 if (P->p_type != PT_GNU_RELRO)
2058 PageAlign(P->FirstSec);
2060 // Find the first section after PT_GNU_RELRO. If it is in a PT_LOAD we
2061 // have to align it to a page.
2062 auto End = OutputSections.end();
2063 auto I = std::find(OutputSections.begin(), End, P->LastSec);
2064 if (I == End || (I + 1) == End)
2067 OutputSection *Cmd = (*(I + 1));
2068 if (needsPtLoad(Cmd))
2073 // Compute an in-file position for a given section. The file offset must be the
2074 // same with its virtual address modulo the page size, so that the loader can
2075 // load executables without any address adjustment.
2076 static uint64_t computeFileOffset(OutputSection *OS, uint64_t Off) {
2077 // File offsets are not significant for .bss sections. By convention, we keep
2078 // section offsets monotonically increasing rather than setting to zero.
2079 if (OS->Type == SHT_NOBITS)
2082 // If the section is not in a PT_LOAD, we just have to align it.
2084 return alignTo(Off, OS->Alignment);
2086 // The first section in a PT_LOAD has to have congruent offset and address
2087 // module the page size.
2088 OutputSection *First = OS->PtLoad->FirstSec;
2090 uint64_t Alignment = std::max<uint64_t>(OS->Alignment, Config->MaxPageSize);
2091 return alignTo(Off, Alignment, OS->Addr);
2094 // If two sections share the same PT_LOAD the file offset is calculated
2095 // using this formula: Off2 = Off1 + (VA2 - VA1).
2096 return First->Offset + OS->Addr - First->Addr;
2099 // Set an in-file position to a given section and returns the end position of
2101 static uint64_t setFileOffset(OutputSection *OS, uint64_t Off) {
2102 Off = computeFileOffset(OS, Off);
2105 if (OS->Type == SHT_NOBITS)
2107 return Off + OS->Size;
2110 template <class ELFT> void Writer<ELFT>::assignFileOffsetsBinary() {
2112 for (OutputSection *Sec : OutputSections)
2113 if (Sec->Flags & SHF_ALLOC)
2114 Off = setFileOffset(Sec, Off);
2115 FileSize = alignTo(Off, Config->Wordsize);
2118 static std::string rangeToString(uint64_t Addr, uint64_t Len) {
2119 return "[0x" + utohexstr(Addr) + ", 0x" + utohexstr(Addr + Len - 1) + "]";
2122 // Assign file offsets to output sections.
2123 template <class ELFT> void Writer<ELFT>::assignFileOffsets() {
2125 Off = setFileOffset(Out::ElfHeader, Off);
2126 Off = setFileOffset(Out::ProgramHeaders, Off);
2128 PhdrEntry *LastRX = nullptr;
2129 for (PhdrEntry *P : Phdrs)
2130 if (P->p_type == PT_LOAD && (P->p_flags & PF_X))
2133 for (OutputSection *Sec : OutputSections) {
2134 Off = setFileOffset(Sec, Off);
2135 if (Script->HasSectionsCommand)
2138 // If this is a last section of the last executable segment and that
2139 // segment is the last loadable segment, align the offset of the
2140 // following section to avoid loading non-segments parts of the file.
2141 if (LastRX && LastRX->LastSec == Sec)
2142 Off = alignTo(Off, Target->PageSize);
2145 SectionHeaderOff = alignTo(Off, Config->Wordsize);
2146 FileSize = SectionHeaderOff + (OutputSections.size() + 1) * sizeof(Elf_Shdr);
2148 // Our logic assumes that sections have rising VA within the same segment.
2149 // With use of linker scripts it is possible to violate this rule and get file
2150 // offset overlaps or overflows. That should never happen with a valid script
2151 // which does not move the location counter backwards and usually scripts do
2152 // not do that. Unfortunately, there are apps in the wild, for example, Linux
2153 // kernel, which control segment distribution explicitly and move the counter
2154 // backwards, so we have to allow doing that to support linking them. We
2155 // perform non-critical checks for overlaps in checkSectionOverlap(), but here
2156 // we want to prevent file size overflows because it would crash the linker.
2157 for (OutputSection *Sec : OutputSections) {
2158 if (Sec->Type == SHT_NOBITS)
2160 if ((Sec->Offset > FileSize) || (Sec->Offset + Sec->Size > FileSize))
2161 error("unable to place section " + Sec->Name + " at file offset " +
2162 rangeToString(Sec->Offset, Sec->Size) +
2163 "; check your linker script for overflows");
2167 // Finalize the program headers. We call this function after we assign
2168 // file offsets and VAs to all sections.
2169 template <class ELFT> void Writer<ELFT>::setPhdrs() {
2170 for (PhdrEntry *P : Phdrs) {
2171 OutputSection *First = P->FirstSec;
2172 OutputSection *Last = P->LastSec;
2175 P->p_filesz = Last->Offset - First->Offset;
2176 if (Last->Type != SHT_NOBITS)
2177 P->p_filesz += Last->Size;
2179 P->p_memsz = Last->Addr + Last->Size - First->Addr;
2180 P->p_offset = First->Offset;
2181 P->p_vaddr = First->Addr;
2184 P->p_paddr = First->getLMA();
2187 if (P->p_type == PT_LOAD) {
2188 P->p_align = std::max<uint64_t>(P->p_align, Config->MaxPageSize);
2189 } else if (P->p_type == PT_GNU_RELRO) {
2191 // The glibc dynamic loader rounds the size down, so we need to round up
2192 // to protect the last page. This is a no-op on FreeBSD which always
2194 P->p_memsz = alignTo(P->p_memsz, Target->PageSize);
2197 if (P->p_type == PT_TLS && P->p_memsz) {
2198 if (!Config->Shared &&
2199 (Config->EMachine == EM_ARM || Config->EMachine == EM_AARCH64)) {
2200 // On ARM/AArch64, reserve extra space (8 words) between the thread
2201 // pointer and an executable's TLS segment by overaligning the segment.
2202 // This reservation is needed for backwards compatibility with Android's
2203 // TCB, which allocates several slots after the thread pointer (e.g.
2204 // TLS_SLOT_STACK_GUARD==5). For simplicity, this overalignment is also
2205 // done on other operating systems.
2206 P->p_align = std::max<uint64_t>(P->p_align, Config->Wordsize * 8);
2209 // The TLS pointer goes after PT_TLS for variant 2 targets. At least glibc
2210 // will align it, so round up the size to make sure the offsets are
2212 P->p_memsz = alignTo(P->p_memsz, P->p_align);
2217 // A helper struct for checkSectionOverlap.
2219 struct SectionOffset {
2225 // Check whether sections overlap for a specific address range (file offsets,
2226 // load and virtual adresses).
2227 static void checkOverlap(StringRef Name, std::vector<SectionOffset> &Sections,
2228 bool IsVirtualAddr) {
2229 llvm::sort(Sections, [=](const SectionOffset &A, const SectionOffset &B) {
2230 return A.Offset < B.Offset;
2233 // Finding overlap is easy given a vector is sorted by start position.
2234 // If an element starts before the end of the previous element, they overlap.
2235 for (size_t I = 1, End = Sections.size(); I < End; ++I) {
2236 SectionOffset A = Sections[I - 1];
2237 SectionOffset B = Sections[I];
2238 if (B.Offset >= A.Offset + A.Sec->Size)
2241 // If both sections are in OVERLAY we allow the overlapping of virtual
2242 // addresses, because it is what OVERLAY was designed for.
2243 if (IsVirtualAddr && A.Sec->InOverlay && B.Sec->InOverlay)
2246 errorOrWarn("section " + A.Sec->Name + " " + Name +
2247 " range overlaps with " + B.Sec->Name + "\n>>> " + A.Sec->Name +
2248 " range is " + rangeToString(A.Offset, A.Sec->Size) + "\n>>> " +
2249 B.Sec->Name + " range is " +
2250 rangeToString(B.Offset, B.Sec->Size));
2254 // Check for overlapping sections and address overflows.
2256 // In this function we check that none of the output sections have overlapping
2257 // file offsets. For SHF_ALLOC sections we also check that the load address
2258 // ranges and the virtual address ranges don't overlap
2259 template <class ELFT> void Writer<ELFT>::checkSections() {
2260 // First, check that section's VAs fit in available address space for target.
2261 for (OutputSection *OS : OutputSections)
2262 if ((OS->Addr + OS->Size < OS->Addr) ||
2263 (!ELFT::Is64Bits && OS->Addr + OS->Size > UINT32_MAX))
2264 errorOrWarn("section " + OS->Name + " at 0x" + utohexstr(OS->Addr) +
2265 " of size 0x" + utohexstr(OS->Size) +
2266 " exceeds available address space");
2268 // Check for overlapping file offsets. In this case we need to skip any
2269 // section marked as SHT_NOBITS. These sections don't actually occupy space in
2270 // the file so Sec->Offset + Sec->Size can overlap with others. If --oformat
2271 // binary is specified only add SHF_ALLOC sections are added to the output
2272 // file so we skip any non-allocated sections in that case.
2273 std::vector<SectionOffset> FileOffs;
2274 for (OutputSection *Sec : OutputSections)
2275 if (Sec->Size > 0 && Sec->Type != SHT_NOBITS &&
2276 (!Config->OFormatBinary || (Sec->Flags & SHF_ALLOC)))
2277 FileOffs.push_back({Sec, Sec->Offset});
2278 checkOverlap("file", FileOffs, false);
2280 // When linking with -r there is no need to check for overlapping virtual/load
2281 // addresses since those addresses will only be assigned when the final
2282 // executable/shared object is created.
2283 if (Config->Relocatable)
2286 // Checking for overlapping virtual and load addresses only needs to take
2287 // into account SHF_ALLOC sections since others will not be loaded.
2288 // Furthermore, we also need to skip SHF_TLS sections since these will be
2289 // mapped to other addresses at runtime and can therefore have overlapping
2290 // ranges in the file.
2291 std::vector<SectionOffset> VMAs;
2292 for (OutputSection *Sec : OutputSections)
2293 if (Sec->Size > 0 && (Sec->Flags & SHF_ALLOC) && !(Sec->Flags & SHF_TLS))
2294 VMAs.push_back({Sec, Sec->Addr});
2295 checkOverlap("virtual address", VMAs, true);
2297 // Finally, check that the load addresses don't overlap. This will usually be
2298 // the same as the virtual addresses but can be different when using a linker
2299 // script with AT().
2300 std::vector<SectionOffset> LMAs;
2301 for (OutputSection *Sec : OutputSections)
2302 if (Sec->Size > 0 && (Sec->Flags & SHF_ALLOC) && !(Sec->Flags & SHF_TLS))
2303 LMAs.push_back({Sec, Sec->getLMA()});
2304 checkOverlap("load address", LMAs, false);
2307 // The entry point address is chosen in the following ways.
2309 // 1. the '-e' entry command-line option;
2310 // 2. the ENTRY(symbol) command in a linker control script;
2311 // 3. the value of the symbol _start, if present;
2312 // 4. the number represented by the entry symbol, if it is a number;
2313 // 5. the address of the first byte of the .text section, if present;
2314 // 6. the address 0.
2315 static uint64_t getEntryAddr() {
2317 if (Symbol *B = Symtab->find(Config->Entry))
2322 if (to_integer(Config->Entry, Addr))
2326 if (OutputSection *Sec = findSection(".text")) {
2327 if (Config->WarnMissingEntry)
2328 warn("cannot find entry symbol " + Config->Entry + "; defaulting to 0x" +
2329 utohexstr(Sec->Addr));
2334 if (Config->WarnMissingEntry)
2335 warn("cannot find entry symbol " + Config->Entry +
2336 "; not setting start address");
2340 static uint16_t getELFType() {
2343 if (Config->Relocatable)
2348 static uint8_t getAbiVersion() {
2349 // MIPS non-PIC executable gets ABI version 1.
2350 if (Config->EMachine == EM_MIPS && getELFType() == ET_EXEC &&
2351 (Config->EFlags & (EF_MIPS_PIC | EF_MIPS_CPIC)) == EF_MIPS_CPIC)
2356 template <class ELFT> void Writer<ELFT>::writeHeader() {
2357 uint8_t *Buf = Buffer->getBufferStart();
2359 // For executable segments, the trap instructions are written before writing
2360 // the header. Setting Elf header bytes to zero ensures that any unused bytes
2361 // in header are zero-cleared, instead of having trap instructions.
2362 memset(Buf, 0, sizeof(Elf_Ehdr));
2363 memcpy(Buf, "\177ELF", 4);
2365 // Write the ELF header.
2366 auto *EHdr = reinterpret_cast<Elf_Ehdr *>(Buf);
2367 EHdr->e_ident[EI_CLASS] = Config->Is64 ? ELFCLASS64 : ELFCLASS32;
2368 EHdr->e_ident[EI_DATA] = Config->IsLE ? ELFDATA2LSB : ELFDATA2MSB;
2369 EHdr->e_ident[EI_VERSION] = EV_CURRENT;
2370 EHdr->e_ident[EI_OSABI] = Config->OSABI;
2371 EHdr->e_ident[EI_ABIVERSION] = getAbiVersion();
2372 EHdr->e_type = getELFType();
2373 EHdr->e_machine = Config->EMachine;
2374 EHdr->e_version = EV_CURRENT;
2375 EHdr->e_entry = getEntryAddr();
2376 EHdr->e_shoff = SectionHeaderOff;
2377 EHdr->e_flags = Config->EFlags;
2378 EHdr->e_ehsize = sizeof(Elf_Ehdr);
2379 EHdr->e_phnum = Phdrs.size();
2380 EHdr->e_shentsize = sizeof(Elf_Shdr);
2382 if (!Config->Relocatable) {
2383 EHdr->e_phoff = sizeof(Elf_Ehdr);
2384 EHdr->e_phentsize = sizeof(Elf_Phdr);
2387 // Write the program header table.
2388 auto *HBuf = reinterpret_cast<Elf_Phdr *>(Buf + EHdr->e_phoff);
2389 for (PhdrEntry *P : Phdrs) {
2390 HBuf->p_type = P->p_type;
2391 HBuf->p_flags = P->p_flags;
2392 HBuf->p_offset = P->p_offset;
2393 HBuf->p_vaddr = P->p_vaddr;
2394 HBuf->p_paddr = P->p_paddr;
2395 HBuf->p_filesz = P->p_filesz;
2396 HBuf->p_memsz = P->p_memsz;
2397 HBuf->p_align = P->p_align;
2401 // Write the section header table.
2403 // The ELF header can only store numbers up to SHN_LORESERVE in the e_shnum
2404 // and e_shstrndx fields. When the value of one of these fields exceeds
2405 // SHN_LORESERVE ELF requires us to put sentinel values in the ELF header and
2406 // use fields in the section header at index 0 to store
2407 // the value. The sentinel values and fields are:
2408 // e_shnum = 0, SHdrs[0].sh_size = number of sections.
2409 // e_shstrndx = SHN_XINDEX, SHdrs[0].sh_link = .shstrtab section index.
2410 auto *SHdrs = reinterpret_cast<Elf_Shdr *>(Buf + EHdr->e_shoff);
2411 size_t Num = OutputSections.size() + 1;
2412 if (Num >= SHN_LORESERVE)
2413 SHdrs->sh_size = Num;
2415 EHdr->e_shnum = Num;
2417 uint32_t StrTabIndex = In.ShStrTab->getParent()->SectionIndex;
2418 if (StrTabIndex >= SHN_LORESERVE) {
2419 SHdrs->sh_link = StrTabIndex;
2420 EHdr->e_shstrndx = SHN_XINDEX;
2422 EHdr->e_shstrndx = StrTabIndex;
2425 for (OutputSection *Sec : OutputSections)
2426 Sec->writeHeaderTo<ELFT>(++SHdrs);
2429 // Open a result file.
2430 template <class ELFT> void Writer<ELFT>::openFile() {
2431 uint64_t MaxSize = Config->Is64 ? INT64_MAX : UINT32_MAX;
2432 if (MaxSize < FileSize) {
2433 error("output file too large: " + Twine(FileSize) + " bytes");
2437 unlinkAsync(Config->OutputFile);
2439 if (!Config->Relocatable)
2440 Flags = FileOutputBuffer::F_executable;
2441 Expected<std::unique_ptr<FileOutputBuffer>> BufferOrErr =
2442 FileOutputBuffer::create(Config->OutputFile, FileSize, Flags);
2445 error("failed to open " + Config->OutputFile + ": " +
2446 llvm::toString(BufferOrErr.takeError()));
2448 Buffer = std::move(*BufferOrErr);
2451 template <class ELFT> void Writer<ELFT>::writeSectionsBinary() {
2452 uint8_t *Buf = Buffer->getBufferStart();
2453 for (OutputSection *Sec : OutputSections)
2454 if (Sec->Flags & SHF_ALLOC)
2455 Sec->writeTo<ELFT>(Buf + Sec->Offset);
2458 static void fillTrap(uint8_t *I, uint8_t *End) {
2459 for (; I + 4 <= End; I += 4)
2460 memcpy(I, &Target->TrapInstr, 4);
2463 // Fill the last page of executable segments with trap instructions
2464 // instead of leaving them as zero. Even though it is not required by any
2465 // standard, it is in general a good thing to do for security reasons.
2467 // We'll leave other pages in segments as-is because the rest will be
2468 // overwritten by output sections.
2469 template <class ELFT> void Writer<ELFT>::writeTrapInstr() {
2470 if (Script->HasSectionsCommand)
2473 // Fill the last page.
2474 uint8_t *Buf = Buffer->getBufferStart();
2475 for (PhdrEntry *P : Phdrs)
2476 if (P->p_type == PT_LOAD && (P->p_flags & PF_X))
2477 fillTrap(Buf + alignDown(P->p_offset + P->p_filesz, Target->PageSize),
2478 Buf + alignTo(P->p_offset + P->p_filesz, Target->PageSize));
2480 // Round up the file size of the last segment to the page boundary iff it is
2481 // an executable segment to ensure that other tools don't accidentally
2482 // trim the instruction padding (e.g. when stripping the file).
2483 PhdrEntry *Last = nullptr;
2484 for (PhdrEntry *P : Phdrs)
2485 if (P->p_type == PT_LOAD)
2488 if (Last && (Last->p_flags & PF_X))
2489 Last->p_memsz = Last->p_filesz = alignTo(Last->p_filesz, Target->PageSize);
2492 // Write section contents to a mmap'ed file.
2493 template <class ELFT> void Writer<ELFT>::writeSections() {
2494 uint8_t *Buf = Buffer->getBufferStart();
2496 OutputSection *EhFrameHdr = nullptr;
2497 if (In.EhFrameHdr && !In.EhFrameHdr->empty())
2498 EhFrameHdr = In.EhFrameHdr->getParent();
2500 // In -r or -emit-relocs mode, write the relocation sections first as in
2501 // ELf_Rel targets we might find out that we need to modify the relocated
2502 // section while doing it.
2503 for (OutputSection *Sec : OutputSections)
2504 if (Sec->Type == SHT_REL || Sec->Type == SHT_RELA)
2505 Sec->writeTo<ELFT>(Buf + Sec->Offset);
2507 for (OutputSection *Sec : OutputSections)
2508 if (Sec != EhFrameHdr && Sec->Type != SHT_REL && Sec->Type != SHT_RELA)
2509 Sec->writeTo<ELFT>(Buf + Sec->Offset);
2511 // The .eh_frame_hdr depends on .eh_frame section contents, therefore
2512 // it should be written after .eh_frame is written.
2514 EhFrameHdr->writeTo<ELFT>(Buf + EhFrameHdr->Offset);
2517 template <class ELFT> void Writer<ELFT>::writeBuildId() {
2518 if (!In.BuildId || !In.BuildId->getParent())
2521 // Compute a hash of all sections of the output file.
2522 uint8_t *Start = Buffer->getBufferStart();
2523 uint8_t *End = Start + FileSize;
2524 In.BuildId->writeBuildId({Start, End});
2527 template void elf::writeResult<ELF32LE>();
2528 template void elf::writeResult<ELF32BE>();
2529 template void elf::writeResult<ELF64LE>();
2530 template void elf::writeResult<ELF64BE>();