1 //===- Writer.cpp ---------------------------------------------------------===//
2 //
3 // The LLVM Linker
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9
10 #include "Writer.h"
11 #include "AArch64ErrataFix.h"
12 #include "CallGraphSort.h"
13 #include "Config.h"
14 #include "Filesystem.h"
15 #include "LinkerScript.h"
16 #include "MapFile.h"
17 #include "OutputSections.h"
18 #include "Relocations.h"
19 #include "SymbolTable.h"
20 #include "Symbols.h"
21 #include "SyntheticSections.h"
22 #include "Target.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"
28 #include <climits>
29
30 using namespace llvm;
31 using namespace llvm::ELF;
32 using namespace llvm::object;
33 using namespace llvm::support;
34 using namespace llvm::support::endian;
35
36 using namespace lld;
37 using namespace lld::elf;
38
39 namespace {
40 // The writer writes a SymbolTable result to a file.
41 template <class ELFT> class Writer {
42 public:
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;
47
48 void run();
49
50 private:
51 void copyLocalSymbols();
52 void addSectionSymbols();
53 void forEachRelSec(llvm::function_ref<void(InputSectionBase &)> Fn);
54 void sortSections();
55 void resolveShfLinkOrder();
56 void sortInputSections();
57 void finalizeSections();
58 void setReservedSymbolSections();
59
60 std::vector<PhdrEntry *> createPhdrs();
61 void removeEmptyPTLoad();
62 void addPtArmExid(std::vector<PhdrEntry *> &Phdrs);
63 void assignFileOffsets();
64 void assignFileOffsetsBinary();
65 void setPhdrs();
66 void checkSections();
67 void fixSectionAlignments();
68 void openFile();
69 void writeTrapInstr();
70 void writeHeader();
71 void writeSections();
72 void writeSectionsBinary();
73 void writeBuildId();
74
75 std::unique_ptr<FileOutputBuffer> &Buffer;
76
77 void addRelIpltSymbols();
78 void addStartEndSymbols();
79 void addStartStopSymbols(OutputSection *Sec);
80 uint64_t getEntryAddr();
81
82 std::vector<PhdrEntry *> Phdrs;
83
84 uint64_t FileSize;
85 uint64_t SectionHeaderOff;
86 };
87 } // anonymous namespace
88
isSectionPrefix(StringRef Prefix,StringRef Name)89 static bool isSectionPrefix(StringRef Prefix, StringRef Name) {
90 return Name.startswith(Prefix) || Name == Prefix.drop_back();
91 }
92
getOutputSectionName(const InputSectionBase * S)93 StringRef elf::getOutputSectionName(const InputSectionBase *S) {
94 if (Config->Relocatable)
95 return S->Name;
96
97 // This is for --emit-relocs. If .text.foo is emitted as .text.bar, we want
98 // to emit .rela.text.foo as .rela.text.bar for consistency (this is not
99 // technically required, but not doing it is odd). This code guarantees that.
100 if (auto *IS = dyn_cast<InputSection>(S)) {
101 if (InputSectionBase *Rel = IS->getRelocatedSection()) {
102 OutputSection *Out = Rel->getOutputSection();
103 if (S->Type == SHT_RELA)
104 return Saver.save(".rela" + Out->Name);
105 return Saver.save(".rel" + Out->Name);
106 }
107 }
108
109 // This check is for -z keep-text-section-prefix. This option separates text
110 // sections with prefix ".text.hot", ".text.unlikely", ".text.startup" or
111 // ".text.exit".
112 // When enabled, this allows identifying the hot code region (.text.hot) in
113 // the final binary which can be selectively mapped to huge pages or mlocked,
114 // for instance.
115 if (Config->ZKeepTextSectionPrefix)
116 for (StringRef V :
117 {".text.hot.", ".text.unlikely.", ".text.startup.", ".text.exit."}) {
118 if (isSectionPrefix(V, S->Name))
119 return V.drop_back();
120 }
121
122 for (StringRef V :
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();
128 }
129
130 // CommonSection is identified as "COMMON" in linker scripts.
131 // By default, it should go to .bss section.
132 if (S->Name == "COMMON")
133 return ".bss";
134
135 return S->Name;
136 }
137
needsInterpSection()138 static bool needsInterpSection() {
139 return !SharedFiles.empty() && !Config->DynamicLinker.empty() &&
140 Script->needsInterpSection();
141 }
142
writeResult()143 template <class ELFT> void elf::writeResult() { Writer<ELFT>().run(); }
144
removeEmptyPTLoad()145 template <class ELFT> void Writer<ELFT>::removeEmptyPTLoad() {
146 llvm::erase_if(Phdrs, [&](const PhdrEntry *P) {
147 if (P->p_type != PT_LOAD)
148 return false;
149 if (!P->FirstSec)
150 return true;
151 uint64_t Size = P->LastSec->Addr + P->LastSec->Size - P->FirstSec->Addr;
152 return Size == 0;
153 });
154 }
155
combineEhFrameSections()156 template <class ELFT> static void combineEhFrameSections() {
157 for (InputSectionBase *&S : InputSections) {
158 EhInputSection *ES = dyn_cast<EhInputSection>(S);
159 if (!ES || !ES->Live)
160 continue;
161
162 InX::EhFrame->addSection<ELFT>(ES);
163 S = nullptr;
164 }
165
166 std::vector<InputSectionBase *> &V = InputSections;
167 V.erase(std::remove(V.begin(), V.end(), nullptr), V.end());
168 }
169
addOptionalRegular(StringRef Name,SectionBase * Sec,uint64_t Val,uint8_t StOther=STV_HIDDEN,uint8_t Binding=STB_GLOBAL)170 static Defined *addOptionalRegular(StringRef Name, SectionBase *Sec,
171 uint64_t Val, uint8_t StOther = STV_HIDDEN,
172 uint8_t Binding = STB_GLOBAL) {
173 Symbol *S = Symtab->find(Name);
174 if (!S || S->isDefined())
175 return nullptr;
176 Symbol *Sym = Symtab->addRegular(Name, StOther, STT_NOTYPE, Val,
177 /*Size=*/0, Binding, Sec,
178 /*File=*/nullptr);
179 return cast<Defined>(Sym);
180 }
181
182 // The linker is expected to define some symbols depending on
183 // the linking result. This function defines such symbols.
addReservedSymbols()184 void elf::addReservedSymbols() {
185 if (Config->EMachine == EM_MIPS) {
186 // Define _gp for MIPS. st_value of _gp symbol will be updated by Writer
187 // so that it points to an absolute address which by default is relative
188 // to GOT. Default offset is 0x7ff0.
189 // See "Global Data Symbols" in Chapter 6 in the following document:
190 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
191 ElfSym::MipsGp = Symtab->addAbsolute("_gp", STV_HIDDEN, STB_GLOBAL);
192
193 // On MIPS O32 ABI, _gp_disp is a magic symbol designates offset between
194 // start of function and 'gp' pointer into GOT.
195 if (Symtab->find("_gp_disp"))
196 ElfSym::MipsGpDisp =
197 Symtab->addAbsolute("_gp_disp", STV_HIDDEN, STB_GLOBAL);
198
199 // The __gnu_local_gp is a magic symbol equal to the current value of 'gp'
200 // pointer. This symbol is used in the code generated by .cpload pseudo-op
201 // in case of using -mno-shared option.
202 // https://sourceware.org/ml/binutils/2004-12/msg00094.html
203 if (Symtab->find("__gnu_local_gp"))
204 ElfSym::MipsLocalGp =
205 Symtab->addAbsolute("__gnu_local_gp", STV_HIDDEN, STB_GLOBAL);
206 }
207
208 // The Power Architecture 64-bit v2 ABI defines a TableOfContents (TOC) which
209 // combines the typical ELF GOT with the small data sections. It commonly
210 // includes .got .toc .sdata .sbss. The .TOC. symbol replaces both
211 // _GLOBAL_OFFSET_TABLE_ and _SDA_BASE_ from the 32-bit ABI. It is used to
212 // represent the TOC base which is offset by 0x8000 bytes from the start of
213 // the .got section.
214 ElfSym::GlobalOffsetTable = addOptionalRegular(
215 (Config->EMachine == EM_PPC64) ? ".TOC." : "_GLOBAL_OFFSET_TABLE_",
216 Out::ElfHeader, Target->GotBaseSymOff);
217
218 // __ehdr_start is the location of ELF file headers. Note that we define
219 // this symbol unconditionally even when using a linker script, which
220 // differs from the behavior implemented by GNU linker which only define
221 // this symbol if ELF headers are in the memory mapped segment.
222 addOptionalRegular("__ehdr_start", Out::ElfHeader, 0, STV_HIDDEN);
223
224 // __executable_start is not documented, but the expectation of at
225 // least the Android libc is that it points to the ELF header.
226 addOptionalRegular("__executable_start", Out::ElfHeader, 0, STV_HIDDEN);
227
228 // __dso_handle symbol is passed to cxa_finalize as a marker to identify
229 // each DSO. The address of the symbol doesn't matter as long as they are
230 // different in different DSOs, so we chose the start address of the DSO.
231 addOptionalRegular("__dso_handle", Out::ElfHeader, 0, STV_HIDDEN);
232
233 // If linker script do layout we do not need to create any standart symbols.
234 if (Script->HasSectionsCommand)
235 return;
236
237 auto Add = [](StringRef S, int64_t Pos) {
238 return addOptionalRegular(S, Out::ElfHeader, Pos, STV_DEFAULT);
239 };
240
241 ElfSym::Bss = Add("__bss_start", 0);
242 ElfSym::End1 = Add("end", -1);
243 ElfSym::End2 = Add("_end", -1);
244 ElfSym::Etext1 = Add("etext", -1);
245 ElfSym::Etext2 = Add("_etext", -1);
246 ElfSym::Edata1 = Add("edata", -1);
247 ElfSym::Edata2 = Add("_edata", -1);
248 }
249
findSection(StringRef Name)250 static OutputSection *findSection(StringRef Name) {
251 for (BaseCommand *Base : Script->SectionCommands)
252 if (auto *Sec = dyn_cast<OutputSection>(Base))
253 if (Sec->Name == Name)
254 return Sec;
255 return nullptr;
256 }
257
258 // Initialize Out members.
createSyntheticSections()259 template <class ELFT> static void createSyntheticSections() {
260 // Initialize all pointers with NULL. This is needed because
261 // you can call lld::elf::main more than once as a library.
262 memset(&Out::First, 0, sizeof(Out));
263
264 auto Add = [](InputSectionBase *Sec) { InputSections.push_back(Sec); };
265
266 InX::DynStrTab = make<StringTableSection>(".dynstr", true);
267 InX::Dynamic = make<DynamicSection<ELFT>>();
268 if (Config->AndroidPackDynRelocs) {
269 InX::RelaDyn = make<AndroidPackedRelocationSection<ELFT>>(
270 Config->IsRela ? ".rela.dyn" : ".rel.dyn");
271 } else {
272 InX::RelaDyn = make<RelocationSection<ELFT>>(
273 Config->IsRela ? ".rela.dyn" : ".rel.dyn", Config->ZCombreloc);
274 }
275 InX::ShStrTab = make<StringTableSection>(".shstrtab", false);
276
277 Out::ProgramHeaders = make<OutputSection>("", 0, SHF_ALLOC);
278 Out::ProgramHeaders->Alignment = Config->Wordsize;
279
280 if (needsInterpSection()) {
281 InX::Interp = createInterpSection();
282 Add(InX::Interp);
283 } else {
284 InX::Interp = nullptr;
285 }
286
287 if (Config->Strip != StripPolicy::All) {
288 InX::StrTab = make<StringTableSection>(".strtab", false);
289 InX::SymTab = make<SymbolTableSection<ELFT>>(*InX::StrTab);
290 InX::SymTabShndx = make<SymtabShndxSection>();
291 }
292
293 if (Config->BuildId != BuildIdKind::None) {
294 InX::BuildId = make<BuildIdSection>();
295 Add(InX::BuildId);
296 }
297
298 InX::Bss = make<BssSection>(".bss", 0, 1);
299 Add(InX::Bss);
300
301 // If there is a SECTIONS command and a .data.rel.ro section name use name
302 // .data.rel.ro.bss so that we match in the .data.rel.ro output section.
303 // This makes sure our relro is contiguous.
304 bool HasDataRelRo = Script->HasSectionsCommand && findSection(".data.rel.ro");
305 InX::BssRelRo =
306 make<BssSection>(HasDataRelRo ? ".data.rel.ro.bss" : ".bss.rel.ro", 0, 1);
307 Add(InX::BssRelRo);
308
309 // Add MIPS-specific sections.
310 if (Config->EMachine == EM_MIPS) {
311 if (!Config->Shared && Config->HasDynSymTab) {
312 InX::MipsRldMap = make<MipsRldMapSection>();
313 Add(InX::MipsRldMap);
314 }
315 if (auto *Sec = MipsAbiFlagsSection<ELFT>::create())
316 Add(Sec);
317 if (auto *Sec = MipsOptionsSection<ELFT>::create())
318 Add(Sec);
319 if (auto *Sec = MipsReginfoSection<ELFT>::create())
320 Add(Sec);
321 }
322
323 if (Config->HasDynSymTab) {
324 InX::DynSymTab = make<SymbolTableSection<ELFT>>(*InX::DynStrTab);
325 Add(InX::DynSymTab);
326
327 In<ELFT>::VerSym = make<VersionTableSection<ELFT>>();
328 Add(In<ELFT>::VerSym);
329
330 if (!Config->VersionDefinitions.empty()) {
331 In<ELFT>::VerDef = make<VersionDefinitionSection<ELFT>>();
332 Add(In<ELFT>::VerDef);
333 }
334
335 In<ELFT>::VerNeed = make<VersionNeedSection<ELFT>>();
336 Add(In<ELFT>::VerNeed);
337
338 if (Config->GnuHash) {
339 InX::GnuHashTab = make<GnuHashTableSection>();
340 Add(InX::GnuHashTab);
341 }
342
343 if (Config->SysvHash) {
344 InX::HashTab = make<HashTableSection>();
345 Add(InX::HashTab);
346 }
347
348 Add(InX::Dynamic);
349 Add(InX::DynStrTab);
350 Add(InX::RelaDyn);
351 }
352
353 if (Config->RelrPackDynRelocs) {
354 InX::RelrDyn = make<RelrSection<ELFT>>();
355 Add(InX::RelrDyn);
356 }
357
358 // Add .got. MIPS' .got is so different from the other archs,
359 // it has its own class.
360 if (Config->EMachine == EM_MIPS) {
361 InX::MipsGot = make<MipsGotSection>();
362 Add(InX::MipsGot);
363 } else {
364 InX::Got = make<GotSection>();
365 Add(InX::Got);
366 }
367
368 InX::GotPlt = make<GotPltSection>();
369 Add(InX::GotPlt);
370 InX::IgotPlt = make<IgotPltSection>();
371 Add(InX::IgotPlt);
372
373 if (Config->GdbIndex) {
374 InX::GdbIndex = GdbIndexSection::create<ELFT>();
375 Add(InX::GdbIndex);
376 }
377
378 // We always need to add rel[a].plt to output if it has entries.
379 // Even for static linking it can contain R_[*]_IRELATIVE relocations.
380 InX::RelaPlt = make<RelocationSection<ELFT>>(
381 Config->IsRela ? ".rela.plt" : ".rel.plt", false /*Sort*/);
382 Add(InX::RelaPlt);
383
384 // The RelaIplt immediately follows .rel.plt (.rel.dyn for ARM) to ensure
385 // that the IRelative relocations are processed last by the dynamic loader.
386 // We cannot place the iplt section in .rel.dyn when Android relocation
387 // packing is enabled because that would cause a section type mismatch.
388 // However, because the Android dynamic loader reads .rel.plt after .rel.dyn,
389 // we can get the desired behaviour by placing the iplt section in .rel.plt.
390 InX::RelaIplt = make<RelocationSection<ELFT>>(
391 (Config->EMachine == EM_ARM && !Config->AndroidPackDynRelocs)
392 ? ".rel.dyn"
393 : InX::RelaPlt->Name,
394 false /*Sort*/);
395 Add(InX::RelaIplt);
396
397 InX::Plt = make<PltSection>(false);
398 Add(InX::Plt);
399 InX::Iplt = make<PltSection>(true);
400 Add(InX::Iplt);
401
402 if (!Config->Relocatable) {
403 if (Config->EhFrameHdr) {
404 InX::EhFrameHdr = make<EhFrameHeader>();
405 Add(InX::EhFrameHdr);
406 }
407 InX::EhFrame = make<EhFrameSection>();
408 Add(InX::EhFrame);
409 }
410
411 if (InX::SymTab)
412 Add(InX::SymTab);
413 if (InX::SymTabShndx)
414 Add(InX::SymTabShndx);
415 Add(InX::ShStrTab);
416 if (InX::StrTab)
417 Add(InX::StrTab);
418
419 if (Config->EMachine == EM_ARM && !Config->Relocatable)
420 // Add a sentinel to terminate .ARM.exidx. It helps an unwinder
421 // to find the exact address range of the last entry.
422 Add(make<ARMExidxSentinelSection>());
423 }
424
425 // The main function of the writer.
run()426 template <class ELFT> void Writer<ELFT>::run() {
427 // Create linker-synthesized sections such as .got or .plt.
428 // Such sections are of type input section.
429 createSyntheticSections<ELFT>();
430
431 if (!Config->Relocatable)
432 combineEhFrameSections<ELFT>();
433
434 // We want to process linker script commands. When SECTIONS command
435 // is given we let it create sections.
436 Script->processSectionCommands();
437
438 // Linker scripts controls how input sections are assigned to output sections.
439 // Input sections that were not handled by scripts are called "orphans", and
440 // they are assigned to output sections by the default rule. Process that.
441 Script->addOrphanSections();
442
443 if (Config->Discard != DiscardPolicy::All)
444 copyLocalSymbols();
445
446 if (Config->CopyRelocs)
447 addSectionSymbols();
448
449 // Now that we have a complete set of output sections. This function
450 // completes section contents. For example, we need to add strings
451 // to the string table, and add entries to .got and .plt.
452 // finalizeSections does that.
453 finalizeSections();
454 if (errorCount())
455 return;
456
457 Script->assignAddresses();
458
459 // If -compressed-debug-sections is specified, we need to compress
460 // .debug_* sections. Do it right now because it changes the size of
461 // output sections.
462 for (OutputSection *Sec : OutputSections)
463 Sec->maybeCompress<ELFT>();
464
465 Script->allocateHeaders(Phdrs);
466
467 // Remove empty PT_LOAD to avoid causing the dynamic linker to try to mmap a
468 // 0 sized region. This has to be done late since only after assignAddresses
469 // we know the size of the sections.
470 removeEmptyPTLoad();
471
472 if (!Config->OFormatBinary)
473 assignFileOffsets();
474 else
475 assignFileOffsetsBinary();
476
477 setPhdrs();
478
479 if (Config->Relocatable) {
480 for (OutputSection *Sec : OutputSections)
481 Sec->Addr = 0;
482 }
483
484 if (Config->CheckSections)
485 checkSections();
486
487 // It does not make sense try to open the file if we have error already.
488 if (errorCount())
489 return;
490 // Write the result down to a file.
491 openFile();
492 if (errorCount())
493 return;
494
495 if (!Config->OFormatBinary) {
496 writeTrapInstr();
497 writeHeader();
498 writeSections();
499 } else {
500 writeSectionsBinary();
501 }
502
503 // Backfill .note.gnu.build-id section content. This is done at last
504 // because the content is usually a hash value of the entire output file.
505 writeBuildId();
506 if (errorCount())
507 return;
508
509 // Handle -Map and -cref options.
510 writeMapFile();
511 writeCrossReferenceTable();
512 if (errorCount())
513 return;
514
515 if (auto E = Buffer->commit())
516 error("failed to write to the output file: " + toString(std::move(E)));
517 }
518
shouldKeepInSymtab(SectionBase * Sec,StringRef SymName,const Symbol & B)519 static bool shouldKeepInSymtab(SectionBase *Sec, StringRef SymName,
520 const Symbol &B) {
521 if (B.isSection())
522 return false;
523
524 if (Config->Discard == DiscardPolicy::None)
525 return true;
526
527 // In ELF assembly .L symbols are normally discarded by the assembler.
528 // If the assembler fails to do so, the linker discards them if
529 // * --discard-locals is used.
530 // * The symbol is in a SHF_MERGE section, which is normally the reason for
531 // the assembler keeping the .L symbol.
532 if (!SymName.startswith(".L") && !SymName.empty())
533 return true;
534
535 if (Config->Discard == DiscardPolicy::Locals)
536 return false;
537
538 return !Sec || !(Sec->Flags & SHF_MERGE);
539 }
540
includeInSymtab(const Symbol & B)541 static bool includeInSymtab(const Symbol &B) {
542 if (!B.isLocal() && !B.IsUsedInRegularObj)
543 return false;
544
545 if (auto *D = dyn_cast<Defined>(&B)) {
546 // Always include absolute symbols.
547 SectionBase *Sec = D->Section;
548 if (!Sec)
549 return true;
550 Sec = Sec->Repl;
551 // Exclude symbols pointing to garbage-collected sections.
552 if (isa<InputSectionBase>(Sec) && !Sec->Live)
553 return false;
554 if (auto *S = dyn_cast<MergeInputSection>(Sec))
555 if (!S->getSectionPiece(D->Value)->Live)
556 return false;
557 return true;
558 }
559 return B.Used;
560 }
561
562 // Local symbols are not in the linker's symbol table. This function scans
563 // each object file's symbol table to copy local symbols to the output.
copyLocalSymbols()564 template <class ELFT> void Writer<ELFT>::copyLocalSymbols() {
565 if (!InX::SymTab)
566 return;
567 for (InputFile *File : ObjectFiles) {
568 ObjFile<ELFT> *F = cast<ObjFile<ELFT>>(File);
569 for (Symbol *B : F->getLocalSymbols()) {
570 if (!B->isLocal())
571 fatal(toString(F) +
572 ": broken object: getLocalSymbols returns a non-local symbol");
573 auto *DR = dyn_cast<Defined>(B);
574
575 // No reason to keep local undefined symbol in symtab.
576 if (!DR)
577 continue;
578 if (!includeInSymtab(*B))
579 continue;
580
581 SectionBase *Sec = DR->Section;
582 if (!shouldKeepInSymtab(Sec, B->getName(), *B))
583 continue;
584 InX::SymTab->addSymbol(B);
585 }
586 }
587 }
588
addSectionSymbols()589 template <class ELFT> void Writer<ELFT>::addSectionSymbols() {
590 // Create a section symbol for each output section so that we can represent
591 // relocations that point to the section. If we know that no relocation is
592 // referring to a section (that happens if the section is a synthetic one), we
593 // don't create a section symbol for that section.
594 for (BaseCommand *Base : Script->SectionCommands) {
595 auto *Sec = dyn_cast<OutputSection>(Base);
596 if (!Sec)
597 continue;
598 auto I = llvm::find_if(Sec->SectionCommands, [](BaseCommand *Base) {
599 if (auto *ISD = dyn_cast<InputSectionDescription>(Base))
600 return !ISD->Sections.empty();
601 return false;
602 });
603 if (I == Sec->SectionCommands.end())
604 continue;
605 InputSection *IS = cast<InputSectionDescription>(*I)->Sections[0];
606
607 // Relocations are not using REL[A] section symbols.
608 if (IS->Type == SHT_REL || IS->Type == SHT_RELA)
609 continue;
610
611 // Unlike other synthetic sections, mergeable output sections contain data
612 // copied from input sections, and there may be a relocation pointing to its
613 // contents if -r or -emit-reloc are given.
614 if (isa<SyntheticSection>(IS) && !(IS->Flags & SHF_MERGE))
615 continue;
616
617 auto *Sym =
618 make<Defined>(IS->File, "", STB_LOCAL, /*StOther=*/0, STT_SECTION,
619 /*Value=*/0, /*Size=*/0, IS);
620 InX::SymTab->addSymbol(Sym);
621 }
622 }
623
624 // Today's loaders have a feature to make segments read-only after
625 // processing dynamic relocations to enhance security. PT_GNU_RELRO
626 // is defined for that.
627 //
628 // This function returns true if a section needs to be put into a
629 // PT_GNU_RELRO segment.
isRelroSection(const OutputSection * Sec)630 static bool isRelroSection(const OutputSection *Sec) {
631 if (!Config->ZRelro)
632 return false;
633
634 uint64_t Flags = Sec->Flags;
635
636 // Non-allocatable or non-writable sections don't need RELRO because
637 // they are not writable or not even mapped to memory in the first place.
638 // RELRO is for sections that are essentially read-only but need to
639 // be writable only at process startup to allow dynamic linker to
640 // apply relocations.
641 if (!(Flags & SHF_ALLOC) || !(Flags & SHF_WRITE))
642 return false;
643
644 // Once initialized, TLS data segments are used as data templates
645 // for a thread-local storage. For each new thread, runtime
646 // allocates memory for a TLS and copy templates there. No thread
647 // are supposed to use templates directly. Thus, it can be in RELRO.
648 if (Flags & SHF_TLS)
649 return true;
650
651 // .init_array, .preinit_array and .fini_array contain pointers to
652 // functions that are executed on process startup or exit. These
653 // pointers are set by the static linker, and they are not expected
654 // to change at runtime. But if you are an attacker, you could do
655 // interesting things by manipulating pointers in .fini_array, for
656 // example. So they are put into RELRO.
657 uint32_t Type = Sec->Type;
658 if (Type == SHT_INIT_ARRAY || Type == SHT_FINI_ARRAY ||
659 Type == SHT_PREINIT_ARRAY)
660 return true;
661
662 // .got contains pointers to external symbols. They are resolved by
663 // the dynamic linker when a module is loaded into memory, and after
664 // that they are not expected to change. So, it can be in RELRO.
665 if (InX::Got && Sec == InX::Got->getParent())
666 return true;
667
668 if (Sec->Name.equals(".toc"))
669 return true;
670
671 // .got.plt contains pointers to external function symbols. They are
672 // by default resolved lazily, so we usually cannot put it into RELRO.
673 // However, if "-z now" is given, the lazy symbol resolution is
674 // disabled, which enables us to put it into RELRO.
675 if (Sec == InX::GotPlt->getParent())
676 return Config->ZNow;
677
678 // .dynamic section contains data for the dynamic linker, and
679 // there's no need to write to it at runtime, so it's better to put
680 // it into RELRO.
681 if (Sec == InX::Dynamic->getParent())
682 return true;
683
684 // Sections with some special names are put into RELRO. This is a
685 // bit unfortunate because section names shouldn't be significant in
686 // ELF in spirit. But in reality many linker features depend on
687 // magic section names.
688 StringRef S = Sec->Name;
689 return S == ".data.rel.ro" || S == ".bss.rel.ro" || S == ".ctors" ||
690 S == ".dtors" || S == ".jcr" || S == ".eh_frame" ||
691 S == ".openbsd.randomdata";
692 }
693
694 // We compute a rank for each section. The rank indicates where the
695 // section should be placed in the file. Instead of using simple
696 // numbers (0,1,2...), we use a series of flags. One for each decision
697 // point when placing the section.
698 // Using flags has two key properties:
699 // * It is easy to check if a give branch was taken.
700 // * It is easy two see how similar two ranks are (see getRankProximity).
701 enum RankFlags {
702 RF_NOT_ADDR_SET = 1 << 18,
703 RF_NOT_INTERP = 1 << 17,
704 RF_NOT_ALLOC = 1 << 16,
705 RF_WRITE = 1 << 15,
706 RF_EXEC_WRITE = 1 << 14,
707 RF_EXEC = 1 << 13,
708 RF_RODATA = 1 << 12,
709 RF_NON_TLS_BSS = 1 << 11,
710 RF_NON_TLS_BSS_RO = 1 << 10,
711 RF_NOT_TLS = 1 << 9,
712 RF_BSS = 1 << 8,
713 RF_NOTE = 1 << 7,
714 RF_PPC_NOT_TOCBSS = 1 << 6,
715 RF_PPC_TOCL = 1 << 5,
716 RF_PPC_TOC = 1 << 4,
717 RF_PPC_GOT = 1 << 3,
718 RF_PPC_BRANCH_LT = 1 << 2,
719 RF_MIPS_GPREL = 1 << 1,
720 RF_MIPS_NOT_GOT = 1 << 0
721 };
722
getSectionRank(const OutputSection * Sec)723 static unsigned getSectionRank(const OutputSection *Sec) {
724 unsigned Rank = 0;
725
726 // We want to put section specified by -T option first, so we
727 // can start assigning VA starting from them later.
728 if (Config->SectionStartMap.count(Sec->Name))
729 return Rank;
730 Rank |= RF_NOT_ADDR_SET;
731
732 // Put .interp first because some loaders want to see that section
733 // on the first page of the executable file when loaded into memory.
734 if (Sec->Name == ".interp")
735 return Rank;
736 Rank |= RF_NOT_INTERP;
737
738 // Allocatable sections go first to reduce the total PT_LOAD size and
739 // so debug info doesn't change addresses in actual code.
740 if (!(Sec->Flags & SHF_ALLOC))
741 return Rank | RF_NOT_ALLOC;
742
743 // Sort sections based on their access permission in the following
744 // order: R, RX, RWX, RW. This order is based on the following
745 // considerations:
746 // * Read-only sections come first such that they go in the
747 // PT_LOAD covering the program headers at the start of the file.
748 // * Read-only, executable sections come next.
749 // * Writable, executable sections follow such that .plt on
750 // architectures where it needs to be writable will be placed
751 // between .text and .data.
752 // * Writable sections come last, such that .bss lands at the very
753 // end of the last PT_LOAD.
754 bool IsExec = Sec->Flags & SHF_EXECINSTR;
755 bool IsWrite = Sec->Flags & SHF_WRITE;
756
757 if (IsExec) {
758 if (IsWrite)
759 Rank |= RF_EXEC_WRITE;
760 else
761 Rank |= RF_EXEC;
762 } else if (IsWrite) {
763 Rank |= RF_WRITE;
764 } else if (Sec->Type == SHT_PROGBITS) {
765 // Make non-executable and non-writable PROGBITS sections (e.g .rodata
766 // .eh_frame) closer to .text. They likely contain PC or GOT relative
767 // relocations and there could be relocation overflow if other huge sections
768 // (.dynstr .dynsym) were placed in between.
769 Rank |= RF_RODATA;
770 }
771
772 // If we got here we know that both A and B are in the same PT_LOAD.
773
774 bool IsTls = Sec->Flags & SHF_TLS;
775 bool IsNoBits = Sec->Type == SHT_NOBITS;
776
777 // The first requirement we have is to put (non-TLS) nobits sections last. The
778 // reason is that the only thing the dynamic linker will see about them is a
779 // p_memsz that is larger than p_filesz. Seeing that it zeros the end of the
780 // PT_LOAD, so that has to correspond to the nobits sections.
781 bool IsNonTlsNoBits = IsNoBits && !IsTls;
782 if (IsNonTlsNoBits)
783 Rank |= RF_NON_TLS_BSS;
784
785 // We place nobits RelRo sections before plain r/w ones, and non-nobits RelRo
786 // sections after r/w ones, so that the RelRo sections are contiguous.
787 bool IsRelRo = isRelroSection(Sec);
788 if (IsNonTlsNoBits && !IsRelRo)
789 Rank |= RF_NON_TLS_BSS_RO;
790 if (!IsNonTlsNoBits && IsRelRo)
791 Rank |= RF_NON_TLS_BSS_RO;
792
793 // The TLS initialization block needs to be a single contiguous block in a R/W
794 // PT_LOAD, so stick TLS sections directly before the other RelRo R/W
795 // sections. The TLS NOBITS sections are placed here as they don't take up
796 // virtual address space in the PT_LOAD.
797 if (!IsTls)
798 Rank |= RF_NOT_TLS;
799
800 // Within the TLS initialization block, the non-nobits sections need to appear
801 // first.
802 if (IsNoBits)
803 Rank |= RF_BSS;
804
805 // We create a NOTE segment for contiguous .note sections, so make
806 // them contigous if there are more than one .note section with the
807 // same attributes.
808 if (Sec->Type == SHT_NOTE)
809 Rank |= RF_NOTE;
810
811 // Some architectures have additional ordering restrictions for sections
812 // within the same PT_LOAD.
813 if (Config->EMachine == EM_PPC64) {
814 // PPC64 has a number of special SHT_PROGBITS+SHF_ALLOC+SHF_WRITE sections
815 // that we would like to make sure appear is a specific order to maximize
816 // their coverage by a single signed 16-bit offset from the TOC base
817 // pointer. Conversely, the special .tocbss section should be first among
818 // all SHT_NOBITS sections. This will put it next to the loaded special
819 // PPC64 sections (and, thus, within reach of the TOC base pointer).
820 StringRef Name = Sec->Name;
821 if (Name != ".tocbss")
822 Rank |= RF_PPC_NOT_TOCBSS;
823
824 if (Name == ".toc1")
825 Rank |= RF_PPC_TOCL;
826
827 if (Name == ".toc")
828 Rank |= RF_PPC_TOC;
829
830 if (Name == ".got")
831 Rank |= RF_PPC_GOT;
832
833 if (Name == ".branch_lt")
834 Rank |= RF_PPC_BRANCH_LT;
835 }
836
837 if (Config->EMachine == EM_MIPS) {
838 // All sections with SHF_MIPS_GPREL flag should be grouped together
839 // because data in these sections is addressable with a gp relative address.
840 if (Sec->Flags & SHF_MIPS_GPREL)
841 Rank |= RF_MIPS_GPREL;
842
843 if (Sec->Name != ".got")
844 Rank |= RF_MIPS_NOT_GOT;
845 }
846
847 return Rank;
848 }
849
compareSections(const BaseCommand * ACmd,const BaseCommand * BCmd)850 static bool compareSections(const BaseCommand *ACmd, const BaseCommand *BCmd) {
851 const OutputSection *A = cast<OutputSection>(ACmd);
852 const OutputSection *B = cast<OutputSection>(BCmd);
853 if (A->SortRank != B->SortRank)
854 return A->SortRank < B->SortRank;
855 if (!(A->SortRank & RF_NOT_ADDR_SET))
856 return Config->SectionStartMap.lookup(A->Name) <
857 Config->SectionStartMap.lookup(B->Name);
858 return false;
859 }
860
add(OutputSection * Sec)861 void PhdrEntry::add(OutputSection *Sec) {
862 LastSec = Sec;
863 if (!FirstSec)
864 FirstSec = Sec;
865 p_align = std::max(p_align, Sec->Alignment);
866 if (p_type == PT_LOAD)
867 Sec->PtLoad = this;
868 }
869
870 // The beginning and the ending of .rel[a].plt section are marked
871 // with __rel[a]_iplt_{start,end} symbols if it is a statically linked
872 // executable. The runtime needs these symbols in order to resolve
873 // all IRELATIVE relocs on startup. For dynamic executables, we don't
874 // need these symbols, since IRELATIVE relocs are resolved through GOT
875 // and PLT. For details, see http://www.airs.com/blog/archives/403.
addRelIpltSymbols()876 template <class ELFT> void Writer<ELFT>::addRelIpltSymbols() {
877 if (Config->Relocatable || needsInterpSection())
878 return;
879 StringRef S = Config->IsRela ? "__rela_iplt_start" : "__rel_iplt_start";
880 addOptionalRegular(S, InX::RelaIplt, 0, STV_HIDDEN, STB_WEAK);
881
882 S = Config->IsRela ? "__rela_iplt_end" : "__rel_iplt_end";
883 ElfSym::RelaIpltEnd =
884 addOptionalRegular(S, InX::RelaIplt, 0, STV_HIDDEN, STB_WEAK);
885 }
886
887 template <class ELFT>
forEachRelSec(llvm::function_ref<void (InputSectionBase &)> Fn)888 void Writer<ELFT>::forEachRelSec(
889 llvm::function_ref<void(InputSectionBase &)> Fn) {
890 // Scan all relocations. Each relocation goes through a series
891 // of tests to determine if it needs special treatment, such as
892 // creating GOT, PLT, copy relocations, etc.
893 // Note that relocations for non-alloc sections are directly
894 // processed by InputSection::relocateNonAlloc.
895 for (InputSectionBase *IS : InputSections)
896 if (IS->Live && isa<InputSection>(IS) && (IS->Flags & SHF_ALLOC))
897 Fn(*IS);
898 for (EhInputSection *ES : InX::EhFrame->Sections)
899 Fn(*ES);
900 }
901
902 // This function generates assignments for predefined symbols (e.g. _end or
903 // _etext) and inserts them into the commands sequence to be processed at the
904 // appropriate time. This ensures that the value is going to be correct by the
905 // time any references to these symbols are processed and is equivalent to
906 // defining these symbols explicitly in the linker script.
setReservedSymbolSections()907 template <class ELFT> void Writer<ELFT>::setReservedSymbolSections() {
908 if (ElfSym::GlobalOffsetTable) {
909 // The _GLOBAL_OFFSET_TABLE_ symbol is defined by target convention usually
910 // to the start of the .got or .got.plt section.
911 InputSection *GotSection = InX::GotPlt;
912 if (!Target->GotBaseSymInGotPlt)
913 GotSection = InX::MipsGot ? cast<InputSection>(InX::MipsGot)
914 : cast<InputSection>(InX::Got);
915 ElfSym::GlobalOffsetTable->Section = GotSection;
916 }
917
918 if (ElfSym::RelaIpltEnd)
919 ElfSym::RelaIpltEnd->Value = InX::RelaIplt->getSize();
920
921 PhdrEntry *Last = nullptr;
922 PhdrEntry *LastRO = nullptr;
923
924 for (PhdrEntry *P : Phdrs) {
925 if (P->p_type != PT_LOAD)
926 continue;
927 Last = P;
928 if (!(P->p_flags & PF_W))
929 LastRO = P;
930 }
931
932 if (LastRO) {
933 // _etext is the first location after the last read-only loadable segment.
934 if (ElfSym::Etext1)
935 ElfSym::Etext1->Section = LastRO->LastSec;
936 if (ElfSym::Etext2)
937 ElfSym::Etext2->Section = LastRO->LastSec;
938 }
939
940 if (Last) {
941 // _edata points to the end of the last mapped initialized section.
942 OutputSection *Edata = nullptr;
943 for (OutputSection *OS : OutputSections) {
944 if (OS->Type != SHT_NOBITS)
945 Edata = OS;
946 if (OS == Last->LastSec)
947 break;
948 }
949
950 if (ElfSym::Edata1)
951 ElfSym::Edata1->Section = Edata;
952 if (ElfSym::Edata2)
953 ElfSym::Edata2->Section = Edata;
954
955 // _end is the first location after the uninitialized data region.
956 if (ElfSym::End1)
957 ElfSym::End1->Section = Last->LastSec;
958 if (ElfSym::End2)
959 ElfSym::End2->Section = Last->LastSec;
960 }
961
962 if (ElfSym::Bss)
963 ElfSym::Bss->Section = findSection(".bss");
964
965 // Setup MIPS _gp_disp/__gnu_local_gp symbols which should
966 // be equal to the _gp symbol's value.
967 if (ElfSym::MipsGp) {
968 // Find GP-relative section with the lowest address
969 // and use this address to calculate default _gp value.
970 for (OutputSection *OS : OutputSections) {
971 if (OS->Flags & SHF_MIPS_GPREL) {
972 ElfSym::MipsGp->Section = OS;
973 ElfSym::MipsGp->Value = 0x7ff0;
974 break;
975 }
976 }
977 }
978 }
979
980 // We want to find how similar two ranks are.
981 // The more branches in getSectionRank that match, the more similar they are.
982 // Since each branch corresponds to a bit flag, we can just use
983 // countLeadingZeros.
getRankProximityAux(OutputSection * A,OutputSection * B)984 static int getRankProximityAux(OutputSection *A, OutputSection *B) {
985 return countLeadingZeros(A->SortRank ^ B->SortRank);
986 }
987
getRankProximity(OutputSection * A,BaseCommand * B)988 static int getRankProximity(OutputSection *A, BaseCommand *B) {
989 if (auto *Sec = dyn_cast<OutputSection>(B))
990 return getRankProximityAux(A, Sec);
991 return -1;
992 }
993
994 // When placing orphan sections, we want to place them after symbol assignments
995 // so that an orphan after
996 // begin_foo = .;
997 // foo : { *(foo) }
998 // end_foo = .;
999 // doesn't break the intended meaning of the begin/end symbols.
1000 // We don't want to go over sections since findOrphanPos is the
1001 // one in charge of deciding the order of the sections.
1002 // We don't want to go over changes to '.', since doing so in
1003 // rx_sec : { *(rx_sec) }
1004 // . = ALIGN(0x1000);
1005 // /* The RW PT_LOAD starts here*/
1006 // rw_sec : { *(rw_sec) }
1007 // would mean that the RW PT_LOAD would become unaligned.
shouldSkip(BaseCommand * Cmd)1008 static bool shouldSkip(BaseCommand *Cmd) {
1009 if (auto *Assign = dyn_cast<SymbolAssignment>(Cmd))
1010 return Assign->Name != ".";
1011 return false;
1012 }
1013
1014 // We want to place orphan sections so that they share as much
1015 // characteristics with their neighbors as possible. For example, if
1016 // both are rw, or both are tls.
1017 template <typename ELFT>
1018 static std::vector<BaseCommand *>::iterator
findOrphanPos(std::vector<BaseCommand * >::iterator B,std::vector<BaseCommand * >::iterator E)1019 findOrphanPos(std::vector<BaseCommand *>::iterator B,
1020 std::vector<BaseCommand *>::iterator E) {
1021 OutputSection *Sec = cast<OutputSection>(*E);
1022
1023 // Find the first element that has as close a rank as possible.
1024 auto I = std::max_element(B, E, [=](BaseCommand *A, BaseCommand *B) {
1025 return getRankProximity(Sec, A) < getRankProximity(Sec, B);
1026 });
1027 if (I == E)
1028 return E;
1029
1030 // Consider all existing sections with the same proximity.
1031 int Proximity = getRankProximity(Sec, *I);
1032 for (; I != E; ++I) {
1033 auto *CurSec = dyn_cast<OutputSection>(*I);
1034 if (!CurSec)
1035 continue;
1036 if (getRankProximity(Sec, CurSec) != Proximity ||
1037 Sec->SortRank < CurSec->SortRank)
1038 break;
1039 }
1040
1041 auto IsOutputSec = [](BaseCommand *Cmd) { return isa<OutputSection>(Cmd); };
1042 auto J = std::find_if(llvm::make_reverse_iterator(I),
1043 llvm::make_reverse_iterator(B), IsOutputSec);
1044 I = J.base();
1045
1046 // As a special case, if the orphan section is the last section, put
1047 // it at the very end, past any other commands.
1048 // This matches bfd's behavior and is convenient when the linker script fully
1049 // specifies the start of the file, but doesn't care about the end (the non
1050 // alloc sections for example).
1051 auto NextSec = std::find_if(I, E, IsOutputSec);
1052 if (NextSec == E)
1053 return E;
1054
1055 while (I != E && shouldSkip(*I))
1056 ++I;
1057 return I;
1058 }
1059
1060 // Builds section order for handling --symbol-ordering-file.
buildSectionOrder()1061 static DenseMap<const InputSectionBase *, int> buildSectionOrder() {
1062 DenseMap<const InputSectionBase *, int> SectionOrder;
1063 // Use the rarely used option -call-graph-ordering-file to sort sections.
1064 if (!Config->CallGraphProfile.empty())
1065 return computeCallGraphProfileOrder();
1066
1067 if (Config->SymbolOrderingFile.empty())
1068 return SectionOrder;
1069
1070 struct SymbolOrderEntry {
1071 int Priority;
1072 bool Present;
1073 };
1074
1075 // Build a map from symbols to their priorities. Symbols that didn't
1076 // appear in the symbol ordering file have the lowest priority 0.
1077 // All explicitly mentioned symbols have negative (higher) priorities.
1078 DenseMap<StringRef, SymbolOrderEntry> SymbolOrder;
1079 int Priority = -Config->SymbolOrderingFile.size();
1080 for (StringRef S : Config->SymbolOrderingFile)
1081 SymbolOrder.insert({S, {Priority++, false}});
1082
1083 // Build a map from sections to their priorities.
1084 auto AddSym = [&](Symbol &Sym) {
1085 auto It = SymbolOrder.find(Sym.getName());
1086 if (It == SymbolOrder.end())
1087 return;
1088 SymbolOrderEntry &Ent = It->second;
1089 Ent.Present = true;
1090
1091 warnUnorderableSymbol(&Sym);
1092
1093 if (auto *D = dyn_cast<Defined>(&Sym)) {
1094 if (auto *Sec = dyn_cast_or_null<InputSectionBase>(D->Section)) {
1095 int &Priority = SectionOrder[cast<InputSectionBase>(Sec->Repl)];
1096 Priority = std::min(Priority, Ent.Priority);
1097 }
1098 }
1099 };
1100 // We want both global and local symbols. We get the global ones from the
1101 // symbol table and iterate the object files for the local ones.
1102 for (Symbol *Sym : Symtab->getSymbols())
1103 if (!Sym->isLazy())
1104 AddSym(*Sym);
1105 for (InputFile *File : ObjectFiles)
1106 for (Symbol *Sym : File->getSymbols())
1107 if (Sym->isLocal())
1108 AddSym(*Sym);
1109
1110 if (Config->WarnSymbolOrdering)
1111 for (auto OrderEntry : SymbolOrder)
1112 if (!OrderEntry.second.Present)
1113 warn("symbol ordering file: no such symbol: " + OrderEntry.first);
1114
1115 return SectionOrder;
1116 }
1117
1118 // Sorts the sections in ISD according to the provided section order.
1119 static void
sortISDBySectionOrder(InputSectionDescription * ISD,const DenseMap<const InputSectionBase *,int> & Order)1120 sortISDBySectionOrder(InputSectionDescription *ISD,
1121 const DenseMap<const InputSectionBase *, int> &Order) {
1122 std::vector<InputSection *> UnorderedSections;
1123 std::vector<std::pair<InputSection *, int>> OrderedSections;
1124 uint64_t UnorderedSize = 0;
1125
1126 for (InputSection *IS : ISD->Sections) {
1127 auto I = Order.find(IS);
1128 if (I == Order.end()) {
1129 UnorderedSections.push_back(IS);
1130 UnorderedSize += IS->getSize();
1131 continue;
1132 }
1133 OrderedSections.push_back({IS, I->second});
1134 }
1135 llvm::sort(
1136 OrderedSections.begin(), OrderedSections.end(),
1137 [&](std::pair<InputSection *, int> A, std::pair<InputSection *, int> B) {
1138 return A.second < B.second;
1139 });
1140
1141 // Find an insertion point for the ordered section list in the unordered
1142 // section list. On targets with limited-range branches, this is the mid-point
1143 // of the unordered section list. This decreases the likelihood that a range
1144 // extension thunk will be needed to enter or exit the ordered region. If the
1145 // ordered section list is a list of hot functions, we can generally expect
1146 // the ordered functions to be called more often than the unordered functions,
1147 // making it more likely that any particular call will be within range, and
1148 // therefore reducing the number of thunks required.
1149 //
1150 // For example, imagine that you have 8MB of hot code and 32MB of cold code.
1151 // If the layout is:
1152 //
1153 // 8MB hot
1154 // 32MB cold
1155 //
1156 // only the first 8-16MB of the cold code (depending on which hot function it
1157 // is actually calling) can call the hot code without a range extension thunk.
1158 // However, if we use this layout:
1159 //
1160 // 16MB cold
1161 // 8MB hot
1162 // 16MB cold
1163 //
1164 // both the last 8-16MB of the first block of cold code and the first 8-16MB
1165 // of the second block of cold code can call the hot code without a thunk. So
1166 // we effectively double the amount of code that could potentially call into
1167 // the hot code without a thunk.
1168 size_t InsPt = 0;
1169 if (Target->ThunkSectionSpacing && !OrderedSections.empty()) {
1170 uint64_t UnorderedPos = 0;
1171 for (; InsPt != UnorderedSections.size(); ++InsPt) {
1172 UnorderedPos += UnorderedSections[InsPt]->getSize();
1173 if (UnorderedPos > UnorderedSize / 2)
1174 break;
1175 }
1176 }
1177
1178 ISD->Sections.clear();
1179 for (InputSection *IS : makeArrayRef(UnorderedSections).slice(0, InsPt))
1180 ISD->Sections.push_back(IS);
1181 for (std::pair<InputSection *, int> P : OrderedSections)
1182 ISD->Sections.push_back(P.first);
1183 for (InputSection *IS : makeArrayRef(UnorderedSections).slice(InsPt))
1184 ISD->Sections.push_back(IS);
1185 }
1186
sortSection(OutputSection * Sec,const DenseMap<const InputSectionBase *,int> & Order)1187 static void sortSection(OutputSection *Sec,
1188 const DenseMap<const InputSectionBase *, int> &Order) {
1189 StringRef Name = Sec->Name;
1190
1191 // Sort input sections by section name suffixes for
1192 // __attribute__((init_priority(N))).
1193 if (Name == ".init_array" || Name == ".fini_array") {
1194 if (!Script->HasSectionsCommand)
1195 Sec->sortInitFini();
1196 return;
1197 }
1198
1199 // Sort input sections by the special rule for .ctors and .dtors.
1200 if (Name == ".ctors" || Name == ".dtors") {
1201 if (!Script->HasSectionsCommand)
1202 Sec->sortCtorsDtors();
1203 return;
1204 }
1205
1206 // Never sort these.
1207 if (Name == ".init" || Name == ".fini")
1208 return;
1209
1210 // Sort input sections by priority using the list provided
1211 // by --symbol-ordering-file.
1212 if (!Order.empty())
1213 for (BaseCommand *B : Sec->SectionCommands)
1214 if (auto *ISD = dyn_cast<InputSectionDescription>(B))
1215 sortISDBySectionOrder(ISD, Order);
1216 }
1217
1218 // If no layout was provided by linker script, we want to apply default
1219 // sorting for special input sections. This also handles --symbol-ordering-file.
sortInputSections()1220 template <class ELFT> void Writer<ELFT>::sortInputSections() {
1221 // Build the order once since it is expensive.
1222 DenseMap<const InputSectionBase *, int> Order = buildSectionOrder();
1223 for (BaseCommand *Base : Script->SectionCommands)
1224 if (auto *Sec = dyn_cast<OutputSection>(Base))
1225 sortSection(Sec, Order);
1226 }
1227
sortSections()1228 template <class ELFT> void Writer<ELFT>::sortSections() {
1229 Script->adjustSectionsBeforeSorting();
1230
1231 // Don't sort if using -r. It is not necessary and we want to preserve the
1232 // relative order for SHF_LINK_ORDER sections.
1233 if (Config->Relocatable)
1234 return;
1235
1236 sortInputSections();
1237
1238 for (BaseCommand *Base : Script->SectionCommands) {
1239 auto *OS = dyn_cast<OutputSection>(Base);
1240 if (!OS)
1241 continue;
1242 OS->SortRank = getSectionRank(OS);
1243
1244 // We want to assign rude approximation values to OutSecOff fields
1245 // to know the relative order of the input sections. We use it for
1246 // sorting SHF_LINK_ORDER sections. See resolveShfLinkOrder().
1247 uint64_t I = 0;
1248 for (InputSection *Sec : getInputSections(OS))
1249 Sec->OutSecOff = I++;
1250 }
1251
1252 if (!Script->HasSectionsCommand) {
1253 // We know that all the OutputSections are contiguous in this case.
1254 auto IsSection = [](BaseCommand *Base) { return isa<OutputSection>(Base); };
1255 std::stable_sort(
1256 llvm::find_if(Script->SectionCommands, IsSection),
1257 llvm::find_if(llvm::reverse(Script->SectionCommands), IsSection).base(),
1258 compareSections);
1259 return;
1260 }
1261
1262 // Orphan sections are sections present in the input files which are
1263 // not explicitly placed into the output file by the linker script.
1264 //
1265 // The sections in the linker script are already in the correct
1266 // order. We have to figuere out where to insert the orphan
1267 // sections.
1268 //
1269 // The order of the sections in the script is arbitrary and may not agree with
1270 // compareSections. This means that we cannot easily define a strict weak
1271 // ordering. To see why, consider a comparison of a section in the script and
1272 // one not in the script. We have a two simple options:
1273 // * Make them equivalent (a is not less than b, and b is not less than a).
1274 // The problem is then that equivalence has to be transitive and we can
1275 // have sections a, b and c with only b in a script and a less than c
1276 // which breaks this property.
1277 // * Use compareSectionsNonScript. Given that the script order doesn't have
1278 // to match, we can end up with sections a, b, c, d where b and c are in the
1279 // script and c is compareSectionsNonScript less than b. In which case d
1280 // can be equivalent to c, a to b and d < a. As a concrete example:
1281 // .a (rx) # not in script
1282 // .b (rx) # in script
1283 // .c (ro) # in script
1284 // .d (ro) # not in script
1285 //
1286 // The way we define an order then is:
1287 // * Sort only the orphan sections. They are in the end right now.
1288 // * Move each orphan section to its preferred position. We try
1289 // to put each section in the last position where it can share
1290 // a PT_LOAD.
1291 //
1292 // There is some ambiguity as to where exactly a new entry should be
1293 // inserted, because Commands contains not only output section
1294 // commands but also other types of commands such as symbol assignment
1295 // expressions. There's no correct answer here due to the lack of the
1296 // formal specification of the linker script. We use heuristics to
1297 // determine whether a new output command should be added before or
1298 // after another commands. For the details, look at shouldSkip
1299 // function.
1300
1301 auto I = Script->SectionCommands.begin();
1302 auto E = Script->SectionCommands.end();
1303 auto NonScriptI = std::find_if(I, E, [](BaseCommand *Base) {
1304 if (auto *Sec = dyn_cast<OutputSection>(Base))
1305 return Sec->SectionIndex == UINT32_MAX;
1306 return false;
1307 });
1308
1309 // Sort the orphan sections.
1310 std::stable_sort(NonScriptI, E, compareSections);
1311
1312 // As a horrible special case, skip the first . assignment if it is before any
1313 // section. We do this because it is common to set a load address by starting
1314 // the script with ". = 0xabcd" and the expectation is that every section is
1315 // after that.
1316 auto FirstSectionOrDotAssignment =
1317 std::find_if(I, E, [](BaseCommand *Cmd) { return !shouldSkip(Cmd); });
1318 if (FirstSectionOrDotAssignment != E &&
1319 isa<SymbolAssignment>(**FirstSectionOrDotAssignment))
1320 ++FirstSectionOrDotAssignment;
1321 I = FirstSectionOrDotAssignment;
1322
1323 while (NonScriptI != E) {
1324 auto Pos = findOrphanPos<ELFT>(I, NonScriptI);
1325 OutputSection *Orphan = cast<OutputSection>(*NonScriptI);
1326
1327 // As an optimization, find all sections with the same sort rank
1328 // and insert them with one rotate.
1329 unsigned Rank = Orphan->SortRank;
1330 auto End = std::find_if(NonScriptI + 1, E, [=](BaseCommand *Cmd) {
1331 return cast<OutputSection>(Cmd)->SortRank != Rank;
1332 });
1333 std::rotate(Pos, NonScriptI, End);
1334 NonScriptI = End;
1335 }
1336
1337 Script->adjustSectionsAfterSorting();
1338 }
1339
compareByFilePosition(InputSection * A,InputSection * B)1340 static bool compareByFilePosition(InputSection *A, InputSection *B) {
1341 // Synthetic, i. e. a sentinel section, should go last.
1342 if (A->kind() == InputSectionBase::Synthetic ||
1343 B->kind() == InputSectionBase::Synthetic)
1344 return A->kind() != InputSectionBase::Synthetic;
1345 InputSection *LA = A->getLinkOrderDep();
1346 InputSection *LB = B->getLinkOrderDep();
1347 OutputSection *AOut = LA->getParent();
1348 OutputSection *BOut = LB->getParent();
1349 if (AOut != BOut)
1350 return AOut->SectionIndex < BOut->SectionIndex;
1351 return LA->OutSecOff < LB->OutSecOff;
1352 }
1353
1354 // This function is used by the --merge-exidx-entries to detect duplicate
1355 // .ARM.exidx sections. It is Arm only.
1356 //
1357 // The .ARM.exidx section is of the form:
1358 // | PREL31 offset to function | Unwind instructions for function |
1359 // where the unwind instructions are either a small number of unwind
1360 // instructions inlined into the table entry, the special CANT_UNWIND value of
1361 // 0x1 or a PREL31 offset into a .ARM.extab Section that contains unwind
1362 // instructions.
1363 //
1364 // We return true if all the unwind instructions in the .ARM.exidx entries of
1365 // Cur can be merged into the last entry of Prev.
isDuplicateArmExidxSec(InputSection * Prev,InputSection * Cur)1366 static bool isDuplicateArmExidxSec(InputSection *Prev, InputSection *Cur) {
1367
1368 // References to .ARM.Extab Sections have bit 31 clear and are not the
1369 // special EXIDX_CANTUNWIND bit-pattern.
1370 auto IsExtabRef = [](uint32_t Unwind) {
1371 return (Unwind & 0x80000000) == 0 && Unwind != 0x1;
1372 };
1373
1374 struct ExidxEntry {
1375 ulittle32_t Fn;
1376 ulittle32_t Unwind;
1377 };
1378
1379 // Get the last table Entry from the previous .ARM.exidx section.
1380 const ExidxEntry &PrevEntry = Prev->getDataAs<ExidxEntry>().back();
1381 if (IsExtabRef(PrevEntry.Unwind))
1382 return false;
1383
1384 // We consider the unwind instructions of an .ARM.exidx table entry
1385 // a duplicate if the previous unwind instructions if:
1386 // - Both are the special EXIDX_CANTUNWIND.
1387 // - Both are the same inline unwind instructions.
1388 // We do not attempt to follow and check links into .ARM.extab tables as
1389 // consecutive identical entries are rare and the effort to check that they
1390 // are identical is high.
1391
1392 for (const ExidxEntry Entry : Cur->getDataAs<ExidxEntry>())
1393 if (IsExtabRef(Entry.Unwind) || Entry.Unwind != PrevEntry.Unwind)
1394 return false;
1395 // All table entries in this .ARM.exidx Section can be merged into the
1396 // previous Section.
1397 return true;
1398 }
1399
resolveShfLinkOrder()1400 template <class ELFT> void Writer<ELFT>::resolveShfLinkOrder() {
1401 for (OutputSection *Sec : OutputSections) {
1402 if (!(Sec->Flags & SHF_LINK_ORDER))
1403 continue;
1404
1405 // Link order may be distributed across several InputSectionDescriptions
1406 // but sort must consider them all at once.
1407 std::vector<InputSection **> ScriptSections;
1408 std::vector<InputSection *> Sections;
1409 for (BaseCommand *Base : Sec->SectionCommands) {
1410 if (auto *ISD = dyn_cast<InputSectionDescription>(Base)) {
1411 for (InputSection *&IS : ISD->Sections) {
1412 ScriptSections.push_back(&IS);
1413 Sections.push_back(IS);
1414 }
1415 }
1416 }
1417 std::stable_sort(Sections.begin(), Sections.end(), compareByFilePosition);
1418
1419 if (!Config->Relocatable && Config->EMachine == EM_ARM &&
1420 Sec->Type == SHT_ARM_EXIDX) {
1421
1422 if (auto *Sentinel = dyn_cast<ARMExidxSentinelSection>(Sections.back())) {
1423 assert(Sections.size() >= 2 &&
1424 "We should create a sentinel section only if there are "
1425 "alive regular exidx sections.");
1426 // The last executable section is required to fill the sentinel.
1427 // Remember it here so that we don't have to find it again.
1428 Sentinel->Highest = Sections[Sections.size() - 2]->getLinkOrderDep();
1429 }
1430
1431 if (Config->MergeArmExidx) {
1432 // The EHABI for the Arm Architecture permits consecutive identical
1433 // table entries to be merged. We use a simple implementation that
1434 // removes a .ARM.exidx Input Section if it can be merged into the
1435 // previous one. This does not require any rewriting of InputSection
1436 // contents but misses opportunities for fine grained deduplication
1437 // where only a subset of the InputSection contents can be merged.
1438 size_t Prev = 0;
1439 // The last one is a sentinel entry which should not be removed.
1440 for (size_t I = 1; I < Sections.size() - 1; ++I) {
1441 if (isDuplicateArmExidxSec(Sections[Prev], Sections[I]))
1442 Sections[I] = nullptr;
1443 else
1444 Prev = I;
1445 }
1446 }
1447 }
1448
1449 for (int I = 0, N = Sections.size(); I < N; ++I)
1450 *ScriptSections[I] = Sections[I];
1451
1452 // Remove the Sections we marked as duplicate earlier.
1453 for (BaseCommand *Base : Sec->SectionCommands)
1454 if (auto *ISD = dyn_cast<InputSectionDescription>(Base))
1455 llvm::erase_if(ISD->Sections, [](InputSection *IS) { return !IS; });
1456 }
1457 }
1458
applySynthetic(const std::vector<SyntheticSection * > & Sections,llvm::function_ref<void (SyntheticSection *)> Fn)1459 static void applySynthetic(const std::vector<SyntheticSection *> &Sections,
1460 llvm::function_ref<void(SyntheticSection *)> Fn) {
1461 for (SyntheticSection *SS : Sections)
1462 if (SS && SS->getParent() && !SS->empty())
1463 Fn(SS);
1464 }
1465
1466 // In order to allow users to manipulate linker-synthesized sections,
1467 // we had to add synthetic sections to the input section list early,
1468 // even before we make decisions whether they are needed. This allows
1469 // users to write scripts like this: ".mygot : { .got }".
1470 //
1471 // Doing it has an unintended side effects. If it turns out that we
1472 // don't need a .got (for example) at all because there's no
1473 // relocation that needs a .got, we don't want to emit .got.
1474 //
1475 // To deal with the above problem, this function is called after
1476 // scanRelocations is called to remove synthetic sections that turn
1477 // out to be empty.
removeUnusedSyntheticSections()1478 static void removeUnusedSyntheticSections() {
1479 // All input synthetic sections that can be empty are placed after
1480 // all regular ones. We iterate over them all and exit at first
1481 // non-synthetic.
1482 for (InputSectionBase *S : llvm::reverse(InputSections)) {
1483 SyntheticSection *SS = dyn_cast<SyntheticSection>(S);
1484 if (!SS)
1485 return;
1486 OutputSection *OS = SS->getParent();
1487 if (!OS || !SS->empty())
1488 continue;
1489
1490 // If we reach here, then SS is an unused synthetic section and we want to
1491 // remove it from corresponding input section description of output section.
1492 for (BaseCommand *B : OS->SectionCommands)
1493 if (auto *ISD = dyn_cast<InputSectionDescription>(B))
1494 llvm::erase_if(ISD->Sections,
1495 [=](InputSection *IS) { return IS == SS; });
1496 }
1497 }
1498
1499 // Returns true if a symbol can be replaced at load-time by a symbol
1500 // with the same name defined in other ELF executable or DSO.
computeIsPreemptible(const Symbol & B)1501 static bool computeIsPreemptible(const Symbol &B) {
1502 assert(!B.isLocal());
1503 // Only symbols that appear in dynsym can be preempted.
1504 if (!B.includeInDynsym())
1505 return false;
1506
1507 // Only default visibility symbols can be preempted.
1508 if (B.Visibility != STV_DEFAULT)
1509 return false;
1510
1511 // At this point copy relocations have not been created yet, so any
1512 // symbol that is not defined locally is preemptible.
1513 if (!B.isDefined())
1514 return true;
1515
1516 // If we have a dynamic list it specifies which local symbols are preemptible.
1517 if (Config->HasDynamicList)
1518 return false;
1519
1520 if (!Config->Shared)
1521 return false;
1522
1523 // -Bsymbolic means that definitions are not preempted.
1524 if (Config->Bsymbolic || (Config->BsymbolicFunctions && B.isFunc()))
1525 return false;
1526 return true;
1527 }
1528
1529 // Create output section objects and add them to OutputSections.
finalizeSections()1530 template <class ELFT> void Writer<ELFT>::finalizeSections() {
1531 Out::DebugInfo = findSection(".debug_info");
1532 Out::PreinitArray = findSection(".preinit_array");
1533 Out::InitArray = findSection(".init_array");
1534 Out::FiniArray = findSection(".fini_array");
1535
1536 // The linker needs to define SECNAME_start, SECNAME_end and SECNAME_stop
1537 // symbols for sections, so that the runtime can get the start and end
1538 // addresses of each section by section name. Add such symbols.
1539 if (!Config->Relocatable) {
1540 addStartEndSymbols();
1541 for (BaseCommand *Base : Script->SectionCommands)
1542 if (auto *Sec = dyn_cast<OutputSection>(Base))
1543 addStartStopSymbols(Sec);
1544 }
1545
1546 // Add _DYNAMIC symbol. Unlike GNU gold, our _DYNAMIC symbol has no type.
1547 // It should be okay as no one seems to care about the type.
1548 // Even the author of gold doesn't remember why gold behaves that way.
1549 // https://sourceware.org/ml/binutils/2002-03/msg00360.html
1550 if (InX::DynSymTab)
1551 Symtab->addRegular("_DYNAMIC", STV_HIDDEN, STT_NOTYPE, 0 /*Value*/,
1552 /*Size=*/0, STB_WEAK, InX::Dynamic,
1553 /*File=*/nullptr);
1554
1555 // Define __rel[a]_iplt_{start,end} symbols if needed.
1556 addRelIpltSymbols();
1557
1558 // This responsible for splitting up .eh_frame section into
1559 // pieces. The relocation scan uses those pieces, so this has to be
1560 // earlier.
1561 applySynthetic({InX::EhFrame},
1562 [](SyntheticSection *SS) { SS->finalizeContents(); });
1563
1564 for (Symbol *S : Symtab->getSymbols()) {
1565 S->IsPreemptible |= computeIsPreemptible(*S);
1566 if (S->isGnuIFunc() && Config->ZIfuncnoplt)
1567 S->ExportDynamic = true;
1568 }
1569
1570 // Scan relocations. This must be done after every symbol is declared so that
1571 // we can correctly decide if a dynamic relocation is needed.
1572 if (!Config->Relocatable)
1573 forEachRelSec(scanRelocations<ELFT>);
1574
1575 if (InX::Plt && !InX::Plt->empty())
1576 InX::Plt->addSymbols();
1577 if (InX::Iplt && !InX::Iplt->empty())
1578 InX::Iplt->addSymbols();
1579
1580 // Now that we have defined all possible global symbols including linker-
1581 // synthesized ones. Visit all symbols to give the finishing touches.
1582 for (Symbol *Sym : Symtab->getSymbols()) {
1583 if (!includeInSymtab(*Sym))
1584 continue;
1585 if (InX::SymTab)
1586 InX::SymTab->addSymbol(Sym);
1587
1588 if (InX::DynSymTab && Sym->includeInDynsym()) {
1589 InX::DynSymTab->addSymbol(Sym);
1590 if (auto *File = dyn_cast_or_null<SharedFile<ELFT>>(Sym->File))
1591 if (File->IsNeeded && !Sym->isUndefined())
1592 In<ELFT>::VerNeed->addSymbol(Sym);
1593 }
1594 }
1595
1596 // Do not proceed if there was an undefined symbol.
1597 if (errorCount())
1598 return;
1599
1600 if (InX::MipsGot)
1601 InX::MipsGot->build<ELFT>();
1602
1603 removeUnusedSyntheticSections();
1604
1605 sortSections();
1606
1607 // Now that we have the final list, create a list of all the
1608 // OutputSections for convenience.
1609 for (BaseCommand *Base : Script->SectionCommands)
1610 if (auto *Sec = dyn_cast<OutputSection>(Base))
1611 OutputSections.push_back(Sec);
1612
1613 // Ensure data sections are not mixed with executable sections when
1614 // -execute-only is used.
1615 if (Config->ExecuteOnly)
1616 for (OutputSection *OS : OutputSections)
1617 if (OS->Flags & SHF_EXECINSTR)
1618 for (InputSection *IS : getInputSections(OS))
1619 if (!(IS->Flags & SHF_EXECINSTR))
1620 error("-execute-only does not support intermingling data and code");
1621
1622 // Prefer command line supplied address over other constraints.
1623 for (OutputSection *Sec : OutputSections) {
1624 auto I = Config->SectionStartMap.find(Sec->Name);
1625 if (I != Config->SectionStartMap.end())
1626 Sec->AddrExpr = [=] { return I->second; };
1627 }
1628
1629 // This is a bit of a hack. A value of 0 means undef, so we set it
1630 // to 1 to make __ehdr_start defined. The section number is not
1631 // particularly relevant.
1632 Out::ElfHeader->SectionIndex = 1;
1633
1634 unsigned I = 1;
1635 for (OutputSection *Sec : OutputSections) {
1636 Sec->SectionIndex = I++;
1637 Sec->ShName = InX::ShStrTab->addString(Sec->Name);
1638 }
1639
1640 // Binary and relocatable output does not have PHDRS.
1641 // The headers have to be created before finalize as that can influence the
1642 // image base and the dynamic section on mips includes the image base.
1643 if (!Config->Relocatable && !Config->OFormatBinary) {
1644 Phdrs = Script->hasPhdrsCommands() ? Script->createPhdrs() : createPhdrs();
1645 addPtArmExid(Phdrs);
1646 Out::ProgramHeaders->Size = sizeof(Elf_Phdr) * Phdrs.size();
1647 }
1648
1649 // Some symbols are defined in term of program headers. Now that we
1650 // have the headers, we can find out which sections they point to.
1651 setReservedSymbolSections();
1652
1653 // Dynamic section must be the last one in this list and dynamic
1654 // symbol table section (DynSymTab) must be the first one.
1655 applySynthetic(
1656 {InX::DynSymTab, InX::Bss, InX::BssRelRo, InX::GnuHashTab,
1657 InX::HashTab, InX::SymTab, InX::SymTabShndx, InX::ShStrTab,
1658 InX::StrTab, In<ELFT>::VerDef, InX::DynStrTab, InX::Got,
1659 InX::MipsGot, InX::IgotPlt, InX::GotPlt, InX::RelaDyn,
1660 InX::RelrDyn, InX::RelaIplt, InX::RelaPlt, InX::Plt,
1661 InX::Iplt, InX::EhFrameHdr, In<ELFT>::VerSym, In<ELFT>::VerNeed,
1662 InX::Dynamic},
1663 [](SyntheticSection *SS) { SS->finalizeContents(); });
1664
1665 if (!Script->HasSectionsCommand && !Config->Relocatable)
1666 fixSectionAlignments();
1667
1668 // After link order processing .ARM.exidx sections can be deduplicated, which
1669 // needs to be resolved before any other address dependent operation.
1670 resolveShfLinkOrder();
1671
1672 // Some architectures need to generate content that depends on the address
1673 // of InputSections. For example some architectures use small displacements
1674 // for jump instructions that is the linker's responsibility for creating
1675 // range extension thunks for. As the generation of the content may also
1676 // alter InputSection addresses we must converge to a fixed point.
1677 if (Target->NeedsThunks || Config->AndroidPackDynRelocs ||
1678 Config->RelrPackDynRelocs) {
1679 ThunkCreator TC;
1680 AArch64Err843419Patcher A64P;
1681 bool Changed;
1682 do {
1683 Script->assignAddresses();
1684 Changed = false;
1685 if (Target->NeedsThunks)
1686 Changed |= TC.createThunks(OutputSections);
1687 if (Config->FixCortexA53Errata843419) {
1688 if (Changed)
1689 Script->assignAddresses();
1690 Changed |= A64P.createFixes();
1691 }
1692 if (InX::MipsGot)
1693 InX::MipsGot->updateAllocSize();
1694 Changed |= InX::RelaDyn->updateAllocSize();
1695 if (InX::RelrDyn)
1696 Changed |= InX::RelrDyn->updateAllocSize();
1697 } while (Changed);
1698 }
1699
1700 // createThunks may have added local symbols to the static symbol table
1701 applySynthetic({InX::SymTab},
1702 [](SyntheticSection *SS) { SS->postThunkContents(); });
1703
1704 // Fill other section headers. The dynamic table is finalized
1705 // at the end because some tags like RELSZ depend on result
1706 // of finalizing other sections.
1707 for (OutputSection *Sec : OutputSections)
1708 Sec->finalize<ELFT>();
1709 }
1710
1711 // The linker is expected to define SECNAME_start and SECNAME_end
1712 // symbols for a few sections. This function defines them.
addStartEndSymbols()1713 template <class ELFT> void Writer<ELFT>::addStartEndSymbols() {
1714 // If a section does not exist, there's ambiguity as to how we
1715 // define _start and _end symbols for an init/fini section. Since
1716 // the loader assume that the symbols are always defined, we need to
1717 // always define them. But what value? The loader iterates over all
1718 // pointers between _start and _end to run global ctors/dtors, so if
1719 // the section is empty, their symbol values don't actually matter
1720 // as long as _start and _end point to the same location.
1721 //
1722 // That said, we don't want to set the symbols to 0 (which is
1723 // probably the simplest value) because that could cause some
1724 // program to fail to link due to relocation overflow, if their
1725 // program text is above 2 GiB. We use the address of the .text
1726 // section instead to prevent that failure.
1727 OutputSection *Default = findSection(".text");
1728 if (!Default)
1729 Default = Out::ElfHeader;
1730 auto Define = [=](StringRef Start, StringRef End, OutputSection *OS) {
1731 if (OS) {
1732 addOptionalRegular(Start, OS, 0);
1733 addOptionalRegular(End, OS, -1);
1734 } else {
1735 addOptionalRegular(Start, Default, 0);
1736 addOptionalRegular(End, Default, 0);
1737 }
1738 };
1739
1740 Define("__preinit_array_start", "__preinit_array_end", Out::PreinitArray);
1741 Define("__init_array_start", "__init_array_end", Out::InitArray);
1742 Define("__fini_array_start", "__fini_array_end", Out::FiniArray);
1743
1744 if (OutputSection *Sec = findSection(".ARM.exidx"))
1745 Define("__exidx_start", "__exidx_end", Sec);
1746 }
1747
1748 // If a section name is valid as a C identifier (which is rare because of
1749 // the leading '.'), linkers are expected to define __start_<secname> and
1750 // __stop_<secname> symbols. They are at beginning and end of the section,
1751 // respectively. This is not requested by the ELF standard, but GNU ld and
1752 // gold provide the feature, and used by many programs.
1753 template <class ELFT>
addStartStopSymbols(OutputSection * Sec)1754 void Writer<ELFT>::addStartStopSymbols(OutputSection *Sec) {
1755 StringRef S = Sec->Name;
1756 if (!isValidCIdentifier(S))
1757 return;
1758 addOptionalRegular(Saver.save("__start_" + S), Sec, 0, STV_PROTECTED);
1759 addOptionalRegular(Saver.save("__stop_" + S), Sec, -1, STV_PROTECTED);
1760 }
1761
needsPtLoad(OutputSection * Sec)1762 static bool needsPtLoad(OutputSection *Sec) {
1763 if (!(Sec->Flags & SHF_ALLOC) || Sec->Noload)
1764 return false;
1765
1766 // Don't allocate VA space for TLS NOBITS sections. The PT_TLS PHDR is
1767 // responsible for allocating space for them, not the PT_LOAD that
1768 // contains the TLS initialization image.
1769 if (Sec->Flags & SHF_TLS && Sec->Type == SHT_NOBITS)
1770 return false;
1771 return true;
1772 }
1773
1774 // Linker scripts are responsible for aligning addresses. Unfortunately, most
1775 // linker scripts are designed for creating two PT_LOADs only, one RX and one
1776 // RW. This means that there is no alignment in the RO to RX transition and we
1777 // cannot create a PT_LOAD there.
computeFlags(uint64_t Flags)1778 static uint64_t computeFlags(uint64_t Flags) {
1779 if (Config->Omagic)
1780 return PF_R | PF_W | PF_X;
1781 if (Config->ExecuteOnly && (Flags & PF_X))
1782 return Flags & ~PF_R;
1783 if (Config->SingleRoRx && !(Flags & PF_W))
1784 return Flags | PF_X;
1785 return Flags;
1786 }
1787
1788 // Decide which program headers to create and which sections to include in each
1789 // one.
createPhdrs()1790 template <class ELFT> std::vector<PhdrEntry *> Writer<ELFT>::createPhdrs() {
1791 std::vector<PhdrEntry *> Ret;
1792 auto AddHdr = [&](unsigned Type, unsigned Flags) -> PhdrEntry * {
1793 Ret.push_back(make<PhdrEntry>(Type, Flags));
1794 return Ret.back();
1795 };
1796
1797 // The first phdr entry is PT_PHDR which describes the program header itself.
1798 AddHdr(PT_PHDR, PF_R)->add(Out::ProgramHeaders);
1799
1800 // PT_INTERP must be the second entry if exists.
1801 if (OutputSection *Cmd = findSection(".interp"))
1802 AddHdr(PT_INTERP, Cmd->getPhdrFlags())->add(Cmd);
1803
1804 // Add the first PT_LOAD segment for regular output sections.
1805 uint64_t Flags = computeFlags(PF_R);
1806 PhdrEntry *Load = AddHdr(PT_LOAD, Flags);
1807
1808 // Add the headers. We will remove them if they don't fit.
1809 Load->add(Out::ElfHeader);
1810 Load->add(Out::ProgramHeaders);
1811
1812 for (OutputSection *Sec : OutputSections) {
1813 if (!(Sec->Flags & SHF_ALLOC))
1814 break;
1815 if (!needsPtLoad(Sec))
1816 continue;
1817
1818 // Segments are contiguous memory regions that has the same attributes
1819 // (e.g. executable or writable). There is one phdr for each segment.
1820 // Therefore, we need to create a new phdr when the next section has
1821 // different flags or is loaded at a discontiguous address or memory
1822 // region using AT or AT> linker script command, respectively. At the same
1823 // time, we don't want to create a separate load segment for the headers,
1824 // even if the first output section has an AT or AT> attribute.
1825 uint64_t NewFlags = computeFlags(Sec->getPhdrFlags());
1826 if (((Sec->LMAExpr ||
1827 (Sec->LMARegion && (Sec->LMARegion != Load->FirstSec->LMARegion))) &&
1828 Load->LastSec != Out::ProgramHeaders) ||
1829 Sec->MemRegion != Load->FirstSec->MemRegion || Flags != NewFlags) {
1830
1831 Load = AddHdr(PT_LOAD, NewFlags);
1832 Flags = NewFlags;
1833 }
1834
1835 Load->add(Sec);
1836 }
1837
1838 // Add a TLS segment if any.
1839 PhdrEntry *TlsHdr = make<PhdrEntry>(PT_TLS, PF_R);
1840 for (OutputSection *Sec : OutputSections)
1841 if (Sec->Flags & SHF_TLS)
1842 TlsHdr->add(Sec);
1843 if (TlsHdr->FirstSec)
1844 Ret.push_back(TlsHdr);
1845
1846 // Add an entry for .dynamic.
1847 if (InX::DynSymTab)
1848 AddHdr(PT_DYNAMIC, InX::Dynamic->getParent()->getPhdrFlags())
1849 ->add(InX::Dynamic->getParent());
1850
1851 // PT_GNU_RELRO includes all sections that should be marked as
1852 // read-only by dynamic linker after proccessing relocations.
1853 // Current dynamic loaders only support one PT_GNU_RELRO PHDR, give
1854 // an error message if more than one PT_GNU_RELRO PHDR is required.
1855 PhdrEntry *RelRo = make<PhdrEntry>(PT_GNU_RELRO, PF_R);
1856 bool InRelroPhdr = false;
1857 bool IsRelroFinished = false;
1858 for (OutputSection *Sec : OutputSections) {
1859 if (!needsPtLoad(Sec))
1860 continue;
1861 if (isRelroSection(Sec)) {
1862 InRelroPhdr = true;
1863 if (!IsRelroFinished)
1864 RelRo->add(Sec);
1865 else
1866 error("section: " + Sec->Name + " is not contiguous with other relro" +
1867 " sections");
1868 } else if (InRelroPhdr) {
1869 InRelroPhdr = false;
1870 IsRelroFinished = true;
1871 }
1872 }
1873 if (RelRo->FirstSec)
1874 Ret.push_back(RelRo);
1875
1876 // PT_GNU_EH_FRAME is a special section pointing on .eh_frame_hdr.
1877 if (!InX::EhFrame->empty() && InX::EhFrameHdr && InX::EhFrame->getParent() &&
1878 InX::EhFrameHdr->getParent())
1879 AddHdr(PT_GNU_EH_FRAME, InX::EhFrameHdr->getParent()->getPhdrFlags())
1880 ->add(InX::EhFrameHdr->getParent());
1881
1882 // PT_OPENBSD_RANDOMIZE is an OpenBSD-specific feature. That makes
1883 // the dynamic linker fill the segment with random data.
1884 if (OutputSection *Cmd = findSection(".openbsd.randomdata"))
1885 AddHdr(PT_OPENBSD_RANDOMIZE, Cmd->getPhdrFlags())->add(Cmd);
1886
1887 // PT_GNU_STACK is a special section to tell the loader to make the
1888 // pages for the stack non-executable. If you really want an executable
1889 // stack, you can pass -z execstack, but that's not recommended for
1890 // security reasons.
1891 unsigned Perm = PF_R | PF_W;
1892 if (Config->ZExecstack)
1893 Perm |= PF_X;
1894 AddHdr(PT_GNU_STACK, Perm)->p_memsz = Config->ZStackSize;
1895
1896 // PT_OPENBSD_WXNEEDED is a OpenBSD-specific header to mark the executable
1897 // is expected to perform W^X violations, such as calling mprotect(2) or
1898 // mmap(2) with PROT_WRITE | PROT_EXEC, which is prohibited by default on
1899 // OpenBSD.
1900 if (Config->ZWxneeded)
1901 AddHdr(PT_OPENBSD_WXNEEDED, PF_X);
1902
1903 // Create one PT_NOTE per a group of contiguous .note sections.
1904 PhdrEntry *Note = nullptr;
1905 for (OutputSection *Sec : OutputSections) {
1906 if (Sec->Type == SHT_NOTE && (Sec->Flags & SHF_ALLOC)) {
1907 if (!Note || Sec->LMAExpr)
1908 Note = AddHdr(PT_NOTE, PF_R);
1909 Note->add(Sec);
1910 } else {
1911 Note = nullptr;
1912 }
1913 }
1914 return Ret;
1915 }
1916
1917 template <class ELFT>
addPtArmExid(std::vector<PhdrEntry * > & Phdrs)1918 void Writer<ELFT>::addPtArmExid(std::vector<PhdrEntry *> &Phdrs) {
1919 if (Config->EMachine != EM_ARM)
1920 return;
1921 auto I = llvm::find_if(OutputSections, [](OutputSection *Cmd) {
1922 return Cmd->Type == SHT_ARM_EXIDX;
1923 });
1924 if (I == OutputSections.end())
1925 return;
1926
1927 // PT_ARM_EXIDX is the ARM EHABI equivalent of PT_GNU_EH_FRAME
1928 PhdrEntry *ARMExidx = make<PhdrEntry>(PT_ARM_EXIDX, PF_R);
1929 ARMExidx->add(*I);
1930 Phdrs.push_back(ARMExidx);
1931 }
1932
1933 // The first section of each PT_LOAD, the first section in PT_GNU_RELRO and the
1934 // first section after PT_GNU_RELRO have to be page aligned so that the dynamic
1935 // linker can set the permissions.
fixSectionAlignments()1936 template <class ELFT> void Writer<ELFT>::fixSectionAlignments() {
1937 auto PageAlign = [](OutputSection *Cmd) {
1938 if (Cmd && !Cmd->AddrExpr)
1939 Cmd->AddrExpr = [=] {
1940 return alignTo(Script->getDot(), Config->MaxPageSize);
1941 };
1942 };
1943
1944 for (const PhdrEntry *P : Phdrs)
1945 if (P->p_type == PT_LOAD && P->FirstSec)
1946 PageAlign(P->FirstSec);
1947
1948 for (const PhdrEntry *P : Phdrs) {
1949 if (P->p_type != PT_GNU_RELRO)
1950 continue;
1951 if (P->FirstSec)
1952 PageAlign(P->FirstSec);
1953 // Find the first section after PT_GNU_RELRO. If it is in a PT_LOAD we
1954 // have to align it to a page.
1955 auto End = OutputSections.end();
1956 auto I = std::find(OutputSections.begin(), End, P->LastSec);
1957 if (I == End || (I + 1) == End)
1958 continue;
1959 OutputSection *Cmd = (*(I + 1));
1960 if (needsPtLoad(Cmd))
1961 PageAlign(Cmd);
1962 }
1963 }
1964
1965 // Adjusts the file alignment for a given output section and returns
1966 // its new file offset. The file offset must be the same with its
1967 // virtual address (modulo the page size) so that the loader can load
1968 // executables without any address adjustment.
getFileAlignment(uint64_t Off,OutputSection * Cmd)1969 static uint64_t getFileAlignment(uint64_t Off, OutputSection *Cmd) {
1970 OutputSection *First = Cmd->PtLoad ? Cmd->PtLoad->FirstSec : nullptr;
1971 // The first section in a PT_LOAD has to have congruent offset and address
1972 // module the page size.
1973 if (Cmd == First)
1974 return alignTo(Off, std::max<uint64_t>(Cmd->Alignment, Config->MaxPageSize),
1975 Cmd->Addr);
1976
1977 // For SHT_NOBITS we don't want the alignment of the section to impact the
1978 // offset of the sections that follow. Since nothing seems to care about the
1979 // sh_offset of the SHT_NOBITS section itself, just ignore it.
1980 if (Cmd->Type == SHT_NOBITS)
1981 return Off;
1982
1983 // If the section is not in a PT_LOAD, we just have to align it.
1984 if (!Cmd->PtLoad)
1985 return alignTo(Off, Cmd->Alignment);
1986
1987 // If two sections share the same PT_LOAD the file offset is calculated
1988 // using this formula: Off2 = Off1 + (VA2 - VA1).
1989 return First->Offset + Cmd->Addr - First->Addr;
1990 }
1991
setOffset(OutputSection * Cmd,uint64_t Off)1992 static uint64_t setOffset(OutputSection *Cmd, uint64_t Off) {
1993 Off = getFileAlignment(Off, Cmd);
1994 Cmd->Offset = Off;
1995
1996 // For SHT_NOBITS we should not count the size.
1997 if (Cmd->Type == SHT_NOBITS)
1998 return Off;
1999
2000 return Off + Cmd->Size;
2001 }
2002
assignFileOffsetsBinary()2003 template <class ELFT> void Writer<ELFT>::assignFileOffsetsBinary() {
2004 uint64_t Off = 0;
2005 for (OutputSection *Sec : OutputSections)
2006 if (Sec->Flags & SHF_ALLOC)
2007 Off = setOffset(Sec, Off);
2008 FileSize = alignTo(Off, Config->Wordsize);
2009 }
2010
rangeToString(uint64_t Addr,uint64_t Len)2011 static std::string rangeToString(uint64_t Addr, uint64_t Len) {
2012 if (Len == 0)
2013 return "<empty range at 0x" + utohexstr(Addr) + ">";
2014 return "[0x" + utohexstr(Addr) + ", 0x" + utohexstr(Addr + Len - 1) + "]";
2015 }
2016
2017 // Assign file offsets to output sections.
assignFileOffsets()2018 template <class ELFT> void Writer<ELFT>::assignFileOffsets() {
2019 uint64_t Off = 0;
2020 Off = setOffset(Out::ElfHeader, Off);
2021 Off = setOffset(Out::ProgramHeaders, Off);
2022
2023 PhdrEntry *LastRX = nullptr;
2024 for (PhdrEntry *P : Phdrs)
2025 if (P->p_type == PT_LOAD && (P->p_flags & PF_X))
2026 LastRX = P;
2027
2028 for (OutputSection *Sec : OutputSections) {
2029 Off = setOffset(Sec, Off);
2030 if (Script->HasSectionsCommand)
2031 continue;
2032 // If this is a last section of the last executable segment and that
2033 // segment is the last loadable segment, align the offset of the
2034 // following section to avoid loading non-segments parts of the file.
2035 if (LastRX && LastRX->LastSec == Sec)
2036 Off = alignTo(Off, Target->PageSize);
2037 }
2038
2039 SectionHeaderOff = alignTo(Off, Config->Wordsize);
2040 FileSize = SectionHeaderOff + (OutputSections.size() + 1) * sizeof(Elf_Shdr);
2041
2042 // Our logic assumes that sections have rising VA within the same segment.
2043 // With use of linker scripts it is possible to violate this rule and get file
2044 // offset overlaps or overflows. That should never happen with a valid script
2045 // which does not move the location counter backwards and usually scripts do
2046 // not do that. Unfortunately, there are apps in the wild, for example, Linux
2047 // kernel, which control segment distribution explicitly and move the counter
2048 // backwards, so we have to allow doing that to support linking them. We
2049 // perform non-critical checks for overlaps in checkSectionOverlap(), but here
2050 // we want to prevent file size overflows because it would crash the linker.
2051 for (OutputSection *Sec : OutputSections) {
2052 if (Sec->Type == SHT_NOBITS)
2053 continue;
2054 if ((Sec->Offset > FileSize) || (Sec->Offset + Sec->Size > FileSize))
2055 error("unable to place section " + Sec->Name + " at file offset " +
2056 rangeToString(Sec->Offset, Sec->Offset + Sec->Size) +
2057 "; check your linker script for overflows");
2058 }
2059 }
2060
2061 // Finalize the program headers. We call this function after we assign
2062 // file offsets and VAs to all sections.
setPhdrs()2063 template <class ELFT> void Writer<ELFT>::setPhdrs() {
2064 for (PhdrEntry *P : Phdrs) {
2065 OutputSection *First = P->FirstSec;
2066 OutputSection *Last = P->LastSec;
2067 if (First) {
2068 P->p_filesz = Last->Offset - First->Offset;
2069 if (Last->Type != SHT_NOBITS)
2070 P->p_filesz += Last->Size;
2071 P->p_memsz = Last->Addr + Last->Size - First->Addr;
2072 P->p_offset = First->Offset;
2073 P->p_vaddr = First->Addr;
2074 if (!P->HasLMA)
2075 P->p_paddr = First->getLMA();
2076 }
2077 if (P->p_type == PT_LOAD)
2078 P->p_align = std::max<uint64_t>(P->p_align, Config->MaxPageSize);
2079 else if (P->p_type == PT_GNU_RELRO) {
2080 P->p_align = 1;
2081 // The glibc dynamic loader rounds the size down, so we need to round up
2082 // to protect the last page. This is a no-op on FreeBSD which always
2083 // rounds up.
2084 P->p_memsz = alignTo(P->p_memsz, Target->PageSize);
2085 }
2086
2087 // The TLS pointer goes after PT_TLS. At least glibc will align it,
2088 // so round up the size to make sure the offsets are correct.
2089 if (P->p_type == PT_TLS) {
2090 Out::TlsPhdr = P;
2091 if (P->p_memsz)
2092 P->p_memsz = alignTo(P->p_memsz, P->p_align);
2093 }
2094 }
2095 }
2096
2097 // A helper struct for checkSectionOverlap.
2098 namespace {
2099 struct SectionOffset {
2100 OutputSection *Sec;
2101 uint64_t Offset;
2102 };
2103 } // namespace
2104
2105 // Check whether sections overlap for a specific address range (file offsets,
2106 // load and virtual adresses).
checkOverlap(StringRef Name,std::vector<SectionOffset> & Sections,bool IsVirtualAddr)2107 static void checkOverlap(StringRef Name, std::vector<SectionOffset> &Sections,
2108 bool IsVirtualAddr) {
2109 llvm::sort(Sections.begin(), Sections.end(),
2110 [=](const SectionOffset &A, const SectionOffset &B) {
2111 return A.Offset < B.Offset;
2112 });
2113
2114 // Finding overlap is easy given a vector is sorted by start position.
2115 // If an element starts before the end of the previous element, they overlap.
2116 for (size_t I = 1, End = Sections.size(); I < End; ++I) {
2117 SectionOffset A = Sections[I - 1];
2118 SectionOffset B = Sections[I];
2119 if (B.Offset >= A.Offset + A.Sec->Size)
2120 continue;
2121
2122 // If both sections are in OVERLAY we allow the overlapping of virtual
2123 // addresses, because it is what OVERLAY was designed for.
2124 if (IsVirtualAddr && A.Sec->InOverlay && B.Sec->InOverlay)
2125 continue;
2126
2127 errorOrWarn("section " + A.Sec->Name + " " + Name +
2128 " range overlaps with " + B.Sec->Name + "\n>>> " + A.Sec->Name +
2129 " range is " + rangeToString(A.Offset, A.Sec->Size) + "\n>>> " +
2130 B.Sec->Name + " range is " +
2131 rangeToString(B.Offset, B.Sec->Size));
2132 }
2133 }
2134
2135 // Check for overlapping sections and address overflows.
2136 //
2137 // In this function we check that none of the output sections have overlapping
2138 // file offsets. For SHF_ALLOC sections we also check that the load address
2139 // ranges and the virtual address ranges don't overlap
checkSections()2140 template <class ELFT> void Writer<ELFT>::checkSections() {
2141 // First, check that section's VAs fit in available address space for target.
2142 for (OutputSection *OS : OutputSections)
2143 if ((OS->Addr + OS->Size < OS->Addr) ||
2144 (!ELFT::Is64Bits && OS->Addr + OS->Size > UINT32_MAX))
2145 errorOrWarn("section " + OS->Name + " at 0x" + utohexstr(OS->Addr) +
2146 " of size 0x" + utohexstr(OS->Size) +
2147 " exceeds available address space");
2148
2149 // Check for overlapping file offsets. In this case we need to skip any
2150 // section marked as SHT_NOBITS. These sections don't actually occupy space in
2151 // the file so Sec->Offset + Sec->Size can overlap with others. If --oformat
2152 // binary is specified only add SHF_ALLOC sections are added to the output
2153 // file so we skip any non-allocated sections in that case.
2154 std::vector<SectionOffset> FileOffs;
2155 for (OutputSection *Sec : OutputSections)
2156 if (0 < Sec->Size && Sec->Type != SHT_NOBITS &&
2157 (!Config->OFormatBinary || (Sec->Flags & SHF_ALLOC)))
2158 FileOffs.push_back({Sec, Sec->Offset});
2159 checkOverlap("file", FileOffs, false);
2160
2161 // When linking with -r there is no need to check for overlapping virtual/load
2162 // addresses since those addresses will only be assigned when the final
2163 // executable/shared object is created.
2164 if (Config->Relocatable)
2165 return;
2166
2167 // Checking for overlapping virtual and load addresses only needs to take
2168 // into account SHF_ALLOC sections since others will not be loaded.
2169 // Furthermore, we also need to skip SHF_TLS sections since these will be
2170 // mapped to other addresses at runtime and can therefore have overlapping
2171 // ranges in the file.
2172 std::vector<SectionOffset> VMAs;
2173 for (OutputSection *Sec : OutputSections)
2174 if (0 < Sec->Size && (Sec->Flags & SHF_ALLOC) && !(Sec->Flags & SHF_TLS))
2175 VMAs.push_back({Sec, Sec->Addr});
2176 checkOverlap("virtual address", VMAs, true);
2177
2178 // Finally, check that the load addresses don't overlap. This will usually be
2179 // the same as the virtual addresses but can be different when using a linker
2180 // script with AT().
2181 std::vector<SectionOffset> LMAs;
2182 for (OutputSection *Sec : OutputSections)
2183 if (0 < Sec->Size && (Sec->Flags & SHF_ALLOC) && !(Sec->Flags & SHF_TLS))
2184 LMAs.push_back({Sec, Sec->getLMA()});
2185 checkOverlap("load address", LMAs, false);
2186 }
2187
2188 // The entry point address is chosen in the following ways.
2189 //
2190 // 1. the '-e' entry command-line option;
2191 // 2. the ENTRY(symbol) command in a linker control script;
2192 // 3. the value of the symbol _start, if present;
2193 // 4. the number represented by the entry symbol, if it is a number;
2194 // 5. the address of the first byte of the .text section, if present;
2195 // 6. the address 0.
getEntryAddr()2196 template <class ELFT> uint64_t Writer<ELFT>::getEntryAddr() {
2197 // Case 1, 2 or 3
2198 if (Symbol *B = Symtab->find(Config->Entry))
2199 return B->getVA();
2200
2201 // Case 4
2202 uint64_t Addr;
2203 if (to_integer(Config->Entry, Addr))
2204 return Addr;
2205
2206 // Case 5
2207 if (OutputSection *Sec = findSection(".text")) {
2208 if (Config->WarnMissingEntry)
2209 warn("cannot find entry symbol " + Config->Entry + "; defaulting to 0x" +
2210 utohexstr(Sec->Addr));
2211 return Sec->Addr;
2212 }
2213
2214 // Case 6
2215 if (Config->WarnMissingEntry)
2216 warn("cannot find entry symbol " + Config->Entry +
2217 "; not setting start address");
2218 return 0;
2219 }
2220
getELFType()2221 static uint16_t getELFType() {
2222 if (Config->Pic)
2223 return ET_DYN;
2224 if (Config->Relocatable)
2225 return ET_REL;
2226 return ET_EXEC;
2227 }
2228
getAbiVersion()2229 static uint8_t getAbiVersion() {
2230 // MIPS non-PIC executable gets ABI version 1.
2231 if (Config->EMachine == EM_MIPS && getELFType() == ET_EXEC &&
2232 (Config->EFlags & (EF_MIPS_PIC | EF_MIPS_CPIC)) == EF_MIPS_CPIC)
2233 return 1;
2234 return 0;
2235 }
2236
writeHeader()2237 template <class ELFT> void Writer<ELFT>::writeHeader() {
2238 uint8_t *Buf = Buffer->getBufferStart();
2239 // For executable segments, the trap instructions are written before writing
2240 // the header. Setting Elf header bytes to zero ensures that any unused bytes
2241 // in header are zero-cleared, instead of having trap instructions.
2242 memset(Buf, 0, sizeof(Elf_Ehdr));
2243 memcpy(Buf, "\177ELF", 4);
2244
2245 // Write the ELF header.
2246 auto *EHdr = reinterpret_cast<Elf_Ehdr *>(Buf);
2247 EHdr->e_ident[EI_CLASS] = Config->Is64 ? ELFCLASS64 : ELFCLASS32;
2248 EHdr->e_ident[EI_DATA] = Config->IsLE ? ELFDATA2LSB : ELFDATA2MSB;
2249 EHdr->e_ident[EI_VERSION] = EV_CURRENT;
2250 EHdr->e_ident[EI_OSABI] = Config->OSABI;
2251 EHdr->e_ident[EI_ABIVERSION] = getAbiVersion();
2252 EHdr->e_type = getELFType();
2253 EHdr->e_machine = Config->EMachine;
2254 EHdr->e_version = EV_CURRENT;
2255 EHdr->e_entry = getEntryAddr();
2256 EHdr->e_shoff = SectionHeaderOff;
2257 EHdr->e_flags = Config->EFlags;
2258 EHdr->e_ehsize = sizeof(Elf_Ehdr);
2259 EHdr->e_phnum = Phdrs.size();
2260 EHdr->e_shentsize = sizeof(Elf_Shdr);
2261
2262 if (!Config->Relocatable) {
2263 EHdr->e_phoff = sizeof(Elf_Ehdr);
2264 EHdr->e_phentsize = sizeof(Elf_Phdr);
2265 }
2266
2267 // Write the program header table.
2268 auto *HBuf = reinterpret_cast<Elf_Phdr *>(Buf + EHdr->e_phoff);
2269 for (PhdrEntry *P : Phdrs) {
2270 HBuf->p_type = P->p_type;
2271 HBuf->p_flags = P->p_flags;
2272 HBuf->p_offset = P->p_offset;
2273 HBuf->p_vaddr = P->p_vaddr;
2274 HBuf->p_paddr = P->p_paddr;
2275 HBuf->p_filesz = P->p_filesz;
2276 HBuf->p_memsz = P->p_memsz;
2277 HBuf->p_align = P->p_align;
2278 ++HBuf;
2279 }
2280
2281 // Write the section header table.
2282 //
2283 // The ELF header can only store numbers up to SHN_LORESERVE in the e_shnum
2284 // and e_shstrndx fields. When the value of one of these fields exceeds
2285 // SHN_LORESERVE ELF requires us to put sentinel values in the ELF header and
2286 // use fields in the section header at index 0 to store
2287 // the value. The sentinel values and fields are:
2288 // e_shnum = 0, SHdrs[0].sh_size = number of sections.
2289 // e_shstrndx = SHN_XINDEX, SHdrs[0].sh_link = .shstrtab section index.
2290 auto *SHdrs = reinterpret_cast<Elf_Shdr *>(Buf + EHdr->e_shoff);
2291 size_t Num = OutputSections.size() + 1;
2292 if (Num >= SHN_LORESERVE)
2293 SHdrs->sh_size = Num;
2294 else
2295 EHdr->e_shnum = Num;
2296
2297 uint32_t StrTabIndex = InX::ShStrTab->getParent()->SectionIndex;
2298 if (StrTabIndex >= SHN_LORESERVE) {
2299 SHdrs->sh_link = StrTabIndex;
2300 EHdr->e_shstrndx = SHN_XINDEX;
2301 } else {
2302 EHdr->e_shstrndx = StrTabIndex;
2303 }
2304
2305 for (OutputSection *Sec : OutputSections)
2306 Sec->writeHeaderTo<ELFT>(++SHdrs);
2307 }
2308
2309 // Open a result file.
openFile()2310 template <class ELFT> void Writer<ELFT>::openFile() {
2311 if (!Config->Is64 && FileSize > UINT32_MAX) {
2312 error("output file too large: " + Twine(FileSize) + " bytes");
2313 return;
2314 }
2315
2316 unlinkAsync(Config->OutputFile);
2317 unsigned Flags = 0;
2318 if (!Config->Relocatable)
2319 Flags = FileOutputBuffer::F_executable;
2320 Expected<std::unique_ptr<FileOutputBuffer>> BufferOrErr =
2321 FileOutputBuffer::create(Config->OutputFile, FileSize, Flags);
2322
2323 if (!BufferOrErr)
2324 error("failed to open " + Config->OutputFile + ": " +
2325 llvm::toString(BufferOrErr.takeError()));
2326 else
2327 Buffer = std::move(*BufferOrErr);
2328 }
2329
writeSectionsBinary()2330 template <class ELFT> void Writer<ELFT>::writeSectionsBinary() {
2331 uint8_t *Buf = Buffer->getBufferStart();
2332 for (OutputSection *Sec : OutputSections)
2333 if (Sec->Flags & SHF_ALLOC)
2334 Sec->writeTo<ELFT>(Buf + Sec->Offset);
2335 }
2336
fillTrap(uint8_t * I,uint8_t * End)2337 static void fillTrap(uint8_t *I, uint8_t *End) {
2338 for (; I + 4 <= End; I += 4)
2339 memcpy(I, &Target->TrapInstr, 4);
2340 }
2341
2342 // Fill the last page of executable segments with trap instructions
2343 // instead of leaving them as zero. Even though it is not required by any
2344 // standard, it is in general a good thing to do for security reasons.
2345 //
2346 // We'll leave other pages in segments as-is because the rest will be
2347 // overwritten by output sections.
writeTrapInstr()2348 template <class ELFT> void Writer<ELFT>::writeTrapInstr() {
2349 if (Script->HasSectionsCommand)
2350 return;
2351
2352 // Fill the last page.
2353 uint8_t *Buf = Buffer->getBufferStart();
2354 for (PhdrEntry *P : Phdrs)
2355 if (P->p_type == PT_LOAD && (P->p_flags & PF_X))
2356 fillTrap(Buf + alignDown(P->p_offset + P->p_filesz, Target->PageSize),
2357 Buf + alignTo(P->p_offset + P->p_filesz, Target->PageSize));
2358
2359 // Round up the file size of the last segment to the page boundary iff it is
2360 // an executable segment to ensure that other tools don't accidentally
2361 // trim the instruction padding (e.g. when stripping the file).
2362 PhdrEntry *Last = nullptr;
2363 for (PhdrEntry *P : Phdrs)
2364 if (P->p_type == PT_LOAD)
2365 Last = P;
2366
2367 if (Last && (Last->p_flags & PF_X))
2368 Last->p_memsz = Last->p_filesz = alignTo(Last->p_filesz, Target->PageSize);
2369 }
2370
2371 // Write section contents to a mmap'ed file.
writeSections()2372 template <class ELFT> void Writer<ELFT>::writeSections() {
2373 uint8_t *Buf = Buffer->getBufferStart();
2374
2375 OutputSection *EhFrameHdr = nullptr;
2376 if (InX::EhFrameHdr && !InX::EhFrameHdr->empty())
2377 EhFrameHdr = InX::EhFrameHdr->getParent();
2378
2379 // In -r or -emit-relocs mode, write the relocation sections first as in
2380 // ELf_Rel targets we might find out that we need to modify the relocated
2381 // section while doing it.
2382 for (OutputSection *Sec : OutputSections)
2383 if (Sec->Type == SHT_REL || Sec->Type == SHT_RELA)
2384 Sec->writeTo<ELFT>(Buf + Sec->Offset);
2385
2386 for (OutputSection *Sec : OutputSections)
2387 if (Sec != EhFrameHdr && Sec->Type != SHT_REL && Sec->Type != SHT_RELA)
2388 Sec->writeTo<ELFT>(Buf + Sec->Offset);
2389
2390 // The .eh_frame_hdr depends on .eh_frame section contents, therefore
2391 // it should be written after .eh_frame is written.
2392 if (EhFrameHdr)
2393 EhFrameHdr->writeTo<ELFT>(Buf + EhFrameHdr->Offset);
2394 }
2395
writeBuildId()2396 template <class ELFT> void Writer<ELFT>::writeBuildId() {
2397 if (!InX::BuildId || !InX::BuildId->getParent())
2398 return;
2399
2400 // Compute a hash of all sections of the output file.
2401 uint8_t *Start = Buffer->getBufferStart();
2402 uint8_t *End = Start + FileSize;
2403 InX::BuildId->writeBuildId({Start, End});
2404 }
2405
2406 template void elf::writeResult<ELF32LE>();
2407 template void elf::writeResult<ELF32BE>();
2408 template void elf::writeResult<ELF64LE>();
2409 template void elf::writeResult<ELF64BE>();
2410