1 //===- Writer.cpp ---------------------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8
9 #include "Writer.h"
10 #include "AArch64ErrataFix.h"
11 #include "ARMErrataFix.h"
12 #include "CallGraphSort.h"
13 #include "Config.h"
14 #include "InputFiles.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/Arrays.h"
24 #include "lld/Common/CommonLinkerContext.h"
25 #include "lld/Common/Filesystem.h"
26 #include "lld/Common/Strings.h"
27 #include "llvm/ADT/StringMap.h"
28 #include "llvm/Support/BLAKE3.h"
29 #include "llvm/Support/Parallel.h"
30 #include "llvm/Support/RandomNumberGenerator.h"
31 #include "llvm/Support/TimeProfiler.h"
32 #include "llvm/Support/xxhash.h"
33 #include <climits>
34
35 #define DEBUG_TYPE "lld"
36
37 using namespace llvm;
38 using namespace llvm::ELF;
39 using namespace llvm::object;
40 using namespace llvm::support;
41 using namespace llvm::support::endian;
42 using namespace lld;
43 using namespace lld::elf;
44
45 namespace {
46 // The writer writes a SymbolTable result to a file.
47 template <class ELFT> class Writer {
48 public:
49 LLVM_ELF_IMPORT_TYPES_ELFT(ELFT)
50
Writer()51 Writer() : buffer(errorHandler().outputBuffer) {}
52
53 void run();
54
55 private:
56 void addSectionSymbols();
57 void sortSections();
58 void resolveShfLinkOrder();
59 void finalizeAddressDependentContent();
60 void optimizeBasicBlockJumps();
61 void sortInputSections();
62 void sortOrphanSections();
63 void finalizeSections();
64 void checkExecuteOnly();
65 void setReservedSymbolSections();
66
67 SmallVector<PhdrEntry *, 0> createPhdrs(Partition &part);
68 void addPhdrForSection(Partition &part, unsigned shType, unsigned pType,
69 unsigned pFlags);
70 void assignFileOffsets();
71 void assignFileOffsetsBinary();
72 void setPhdrs(Partition &part);
73 void checkSections();
74 void fixSectionAlignments();
75 void openFile();
76 void writeTrapInstr();
77 void writeHeader();
78 void writeSections();
79 void writeSectionsBinary();
80 void writeBuildId();
81
82 std::unique_ptr<FileOutputBuffer> &buffer;
83
84 void addRelIpltSymbols();
85 void addStartEndSymbols();
86 void addStartStopSymbols(OutputSection &osec);
87
88 uint64_t fileSize;
89 uint64_t sectionHeaderOff;
90 };
91 } // anonymous namespace
92
needsInterpSection()93 static bool needsInterpSection() {
94 return !config->relocatable && !config->shared &&
95 !config->dynamicLinker.empty() && script->needsInterpSection();
96 }
97
writeResult()98 template <class ELFT> void elf::writeResult() {
99 Writer<ELFT>().run();
100 }
101
removeEmptyPTLoad(SmallVector<PhdrEntry *,0> & phdrs)102 static void removeEmptyPTLoad(SmallVector<PhdrEntry *, 0> &phdrs) {
103 auto it = std::stable_partition(
104 phdrs.begin(), phdrs.end(), [&](const PhdrEntry *p) {
105 if (p->p_type != PT_LOAD)
106 return true;
107 if (!p->firstSec)
108 return false;
109 uint64_t size = p->lastSec->addr + p->lastSec->size - p->firstSec->addr;
110 return size != 0;
111 });
112
113 // Clear OutputSection::ptLoad for sections contained in removed
114 // segments.
115 DenseSet<PhdrEntry *> removed(it, phdrs.end());
116 for (OutputSection *sec : outputSections)
117 if (removed.count(sec->ptLoad))
118 sec->ptLoad = nullptr;
119 phdrs.erase(it, phdrs.end());
120 }
121
copySectionsIntoPartitions()122 void elf::copySectionsIntoPartitions() {
123 SmallVector<InputSectionBase *, 0> newSections;
124 const size_t ehSize = ctx.ehInputSections.size();
125 for (unsigned part = 2; part != partitions.size() + 1; ++part) {
126 for (InputSectionBase *s : ctx.inputSections) {
127 if (!(s->flags & SHF_ALLOC) || !s->isLive() || s->type != SHT_NOTE)
128 continue;
129 auto *copy = make<InputSection>(cast<InputSection>(*s));
130 copy->partition = part;
131 newSections.push_back(copy);
132 }
133 for (size_t i = 0; i != ehSize; ++i) {
134 assert(ctx.ehInputSections[i]->isLive());
135 auto *copy = make<EhInputSection>(*ctx.ehInputSections[i]);
136 copy->partition = part;
137 ctx.ehInputSections.push_back(copy);
138 }
139 }
140
141 ctx.inputSections.insert(ctx.inputSections.end(), newSections.begin(),
142 newSections.end());
143 }
144
addOptionalRegular(StringRef name,SectionBase * sec,uint64_t val,uint8_t stOther=STV_HIDDEN)145 static Defined *addOptionalRegular(StringRef name, SectionBase *sec,
146 uint64_t val, uint8_t stOther = STV_HIDDEN) {
147 Symbol *s = symtab.find(name);
148 if (!s || s->isDefined() || s->isCommon())
149 return nullptr;
150
151 s->resolve(Defined{ctx.internalFile, StringRef(), STB_GLOBAL, stOther,
152 STT_NOTYPE, val,
153 /*size=*/0, sec});
154 s->isUsedInRegularObj = true;
155 return cast<Defined>(s);
156 }
157
158 // The linker is expected to define some symbols depending on
159 // the linking result. This function defines such symbols.
addReservedSymbols()160 void elf::addReservedSymbols() {
161 if (config->emachine == EM_MIPS) {
162 auto addAbsolute = [](StringRef name) {
163 Symbol *sym =
164 symtab.addSymbol(Defined{ctx.internalFile, name, STB_GLOBAL,
165 STV_HIDDEN, STT_NOTYPE, 0, 0, nullptr});
166 sym->isUsedInRegularObj = true;
167 return cast<Defined>(sym);
168 };
169 // Define _gp for MIPS. st_value of _gp symbol will be updated by Writer
170 // so that it points to an absolute address which by default is relative
171 // to GOT. Default offset is 0x7ff0.
172 // See "Global Data Symbols" in Chapter 6 in the following document:
173 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
174 ElfSym::mipsGp = addAbsolute("_gp");
175
176 // On MIPS O32 ABI, _gp_disp is a magic symbol designates offset between
177 // start of function and 'gp' pointer into GOT.
178 if (symtab.find("_gp_disp"))
179 ElfSym::mipsGpDisp = addAbsolute("_gp_disp");
180
181 // The __gnu_local_gp is a magic symbol equal to the current value of 'gp'
182 // pointer. This symbol is used in the code generated by .cpload pseudo-op
183 // in case of using -mno-shared option.
184 // https://sourceware.org/ml/binutils/2004-12/msg00094.html
185 if (symtab.find("__gnu_local_gp"))
186 ElfSym::mipsLocalGp = addAbsolute("__gnu_local_gp");
187 } else if (config->emachine == EM_PPC) {
188 // glibc *crt1.o has a undefined reference to _SDA_BASE_. Since we don't
189 // support Small Data Area, define it arbitrarily as 0.
190 addOptionalRegular("_SDA_BASE_", nullptr, 0, STV_HIDDEN);
191 } else if (config->emachine == EM_PPC64) {
192 addPPC64SaveRestore();
193 }
194
195 // The Power Architecture 64-bit v2 ABI defines a TableOfContents (TOC) which
196 // combines the typical ELF GOT with the small data sections. It commonly
197 // includes .got .toc .sdata .sbss. The .TOC. symbol replaces both
198 // _GLOBAL_OFFSET_TABLE_ and _SDA_BASE_ from the 32-bit ABI. It is used to
199 // represent the TOC base which is offset by 0x8000 bytes from the start of
200 // the .got section.
201 // We do not allow _GLOBAL_OFFSET_TABLE_ to be defined by input objects as the
202 // correctness of some relocations depends on its value.
203 StringRef gotSymName =
204 (config->emachine == EM_PPC64) ? ".TOC." : "_GLOBAL_OFFSET_TABLE_";
205
206 if (Symbol *s = symtab.find(gotSymName)) {
207 if (s->isDefined()) {
208 error(toString(s->file) + " cannot redefine linker defined symbol '" +
209 gotSymName + "'");
210 return;
211 }
212
213 uint64_t gotOff = 0;
214 if (config->emachine == EM_PPC64)
215 gotOff = 0x8000;
216
217 s->resolve(Defined{ctx.internalFile, StringRef(), STB_GLOBAL, STV_HIDDEN,
218 STT_NOTYPE, gotOff, /*size=*/0, Out::elfHeader});
219 ElfSym::globalOffsetTable = cast<Defined>(s);
220 }
221
222 // __ehdr_start is the location of ELF file headers. Note that we define
223 // this symbol unconditionally even when using a linker script, which
224 // differs from the behavior implemented by GNU linker which only define
225 // this symbol if ELF headers are in the memory mapped segment.
226 addOptionalRegular("__ehdr_start", Out::elfHeader, 0, STV_HIDDEN);
227
228 // __executable_start is not documented, but the expectation of at
229 // least the Android libc is that it points to the ELF header.
230 addOptionalRegular("__executable_start", Out::elfHeader, 0, STV_HIDDEN);
231
232 // __dso_handle symbol is passed to cxa_finalize as a marker to identify
233 // each DSO. The address of the symbol doesn't matter as long as they are
234 // different in different DSOs, so we chose the start address of the DSO.
235 addOptionalRegular("__dso_handle", Out::elfHeader, 0, STV_HIDDEN);
236
237 // If linker script do layout we do not need to create any standard symbols.
238 if (script->hasSectionsCommand)
239 return;
240
241 auto add = [](StringRef s, int64_t pos) {
242 return addOptionalRegular(s, Out::elfHeader, pos, STV_DEFAULT);
243 };
244
245 ElfSym::bss = add("__bss_start", 0);
246 ElfSym::end1 = add("end", -1);
247 ElfSym::end2 = add("_end", -1);
248 ElfSym::etext1 = add("etext", -1);
249 ElfSym::etext2 = add("_etext", -1);
250 ElfSym::edata1 = add("edata", -1);
251 ElfSym::edata2 = add("_edata", -1);
252 }
253
demoteDefined(Defined & sym,DenseMap<SectionBase *,size_t> & map)254 static void demoteDefined(Defined &sym, DenseMap<SectionBase *, size_t> &map) {
255 if (map.empty())
256 for (auto [i, sec] : llvm::enumerate(sym.file->getSections()))
257 map.try_emplace(sec, i);
258 // Change WEAK to GLOBAL so that if a scanned relocation references sym,
259 // maybeReportUndefined will report an error.
260 uint8_t binding = sym.isWeak() ? uint8_t(STB_GLOBAL) : sym.binding;
261 Undefined(sym.file, sym.getName(), binding, sym.stOther, sym.type,
262 /*discardedSecIdx=*/map.lookup(sym.section))
263 .overwrite(sym);
264 // Eliminate from the symbol table, otherwise we would leave an undefined
265 // symbol if the symbol is unreferenced in the absence of GC.
266 sym.isUsedInRegularObj = false;
267 }
268
269 // If all references to a DSO happen to be weak, the DSO is not added to
270 // DT_NEEDED. If that happens, replace ShardSymbol with Undefined to avoid
271 // dangling references to an unneeded DSO. Use a weak binding to avoid
272 // --no-allow-shlib-undefined diagnostics. Similarly, demote lazy symbols.
273 //
274 // In addition, demote symbols defined in discarded sections, so that
275 // references to /DISCARD/ discarded symbols will lead to errors.
demoteSymbolsAndComputeIsPreemptible()276 static void demoteSymbolsAndComputeIsPreemptible() {
277 llvm::TimeTraceScope timeScope("Demote symbols");
278 DenseMap<InputFile *, DenseMap<SectionBase *, size_t>> sectionIndexMap;
279 for (Symbol *sym : symtab.getSymbols()) {
280 if (auto *d = dyn_cast<Defined>(sym)) {
281 if (d->section && !d->section->isLive())
282 demoteDefined(*d, sectionIndexMap[d->file]);
283 } else {
284 auto *s = dyn_cast<SharedSymbol>(sym);
285 if (sym->isLazy() || (s && !cast<SharedFile>(s->file)->isNeeded)) {
286 uint8_t binding = sym->isLazy() ? sym->binding : uint8_t(STB_WEAK);
287 Undefined(ctx.internalFile, sym->getName(), binding, sym->stOther,
288 sym->type)
289 .overwrite(*sym);
290 sym->versionId = VER_NDX_GLOBAL;
291 }
292 }
293
294 if (config->hasDynSymTab)
295 sym->isPreemptible = computeIsPreemptible(*sym);
296 }
297 }
298
hasMemtag()299 bool elf::hasMemtag() {
300 return config->emachine == EM_AARCH64 &&
301 config->androidMemtagMode != ELF::NT_MEMTAG_LEVEL_NONE;
302 }
303
304 // Fully static executables don't support MTE globals at this point in time, as
305 // we currently rely on:
306 // - A dynamic loader to process relocations, and
307 // - Dynamic entries.
308 // This restriction could be removed in future by re-using some of the ideas
309 // that ifuncs use in fully static executables.
canHaveMemtagGlobals()310 bool elf::canHaveMemtagGlobals() {
311 return hasMemtag() &&
312 (config->relocatable || config->shared || needsInterpSection());
313 }
314
findSection(StringRef name,unsigned partition=1)315 static OutputSection *findSection(StringRef name, unsigned partition = 1) {
316 for (SectionCommand *cmd : script->sectionCommands)
317 if (auto *osd = dyn_cast<OutputDesc>(cmd))
318 if (osd->osec.name == name && osd->osec.partition == partition)
319 return &osd->osec;
320 return nullptr;
321 }
322
createSyntheticSections()323 template <class ELFT> void elf::createSyntheticSections() {
324 // Initialize all pointers with NULL. This is needed because
325 // you can call lld::elf::main more than once as a library.
326 Out::tlsPhdr = nullptr;
327 Out::preinitArray = nullptr;
328 Out::initArray = nullptr;
329 Out::finiArray = nullptr;
330
331 // Add the .interp section first because it is not a SyntheticSection.
332 // The removeUnusedSyntheticSections() function relies on the
333 // SyntheticSections coming last.
334 if (needsInterpSection()) {
335 for (size_t i = 1; i <= partitions.size(); ++i) {
336 InputSection *sec = createInterpSection();
337 sec->partition = i;
338 ctx.inputSections.push_back(sec);
339 }
340 }
341
342 auto add = [](SyntheticSection &sec) { ctx.inputSections.push_back(&sec); };
343
344 in.shStrTab = std::make_unique<StringTableSection>(".shstrtab", false);
345
346 Out::programHeaders = make<OutputSection>("", 0, SHF_ALLOC);
347 Out::programHeaders->addralign = config->wordsize;
348
349 if (config->strip != StripPolicy::All) {
350 in.strTab = std::make_unique<StringTableSection>(".strtab", false);
351 in.symTab = std::make_unique<SymbolTableSection<ELFT>>(*in.strTab);
352 in.symTabShndx = std::make_unique<SymtabShndxSection>();
353 }
354
355 in.bss = std::make_unique<BssSection>(".bss", 0, 1);
356 add(*in.bss);
357
358 // If there is a SECTIONS command and a .data.rel.ro section name use name
359 // .data.rel.ro.bss so that we match in the .data.rel.ro output section.
360 // This makes sure our relro is contiguous.
361 bool hasDataRelRo = script->hasSectionsCommand && findSection(".data.rel.ro");
362 in.bssRelRo = std::make_unique<BssSection>(
363 hasDataRelRo ? ".data.rel.ro.bss" : ".bss.rel.ro", 0, 1);
364 add(*in.bssRelRo);
365
366 // Add MIPS-specific sections.
367 if (config->emachine == EM_MIPS) {
368 if (!config->shared && config->hasDynSymTab) {
369 in.mipsRldMap = std::make_unique<MipsRldMapSection>();
370 add(*in.mipsRldMap);
371 }
372 if ((in.mipsAbiFlags = MipsAbiFlagsSection<ELFT>::create()))
373 add(*in.mipsAbiFlags);
374 if ((in.mipsOptions = MipsOptionsSection<ELFT>::create()))
375 add(*in.mipsOptions);
376 if ((in.mipsReginfo = MipsReginfoSection<ELFT>::create()))
377 add(*in.mipsReginfo);
378 }
379
380 StringRef relaDynName = config->isRela ? ".rela.dyn" : ".rel.dyn";
381
382 const unsigned threadCount = config->threadCount;
383 for (Partition &part : partitions) {
384 auto add = [&](SyntheticSection &sec) {
385 sec.partition = part.getNumber();
386 ctx.inputSections.push_back(&sec);
387 };
388
389 if (!part.name.empty()) {
390 part.elfHeader = std::make_unique<PartitionElfHeaderSection<ELFT>>();
391 part.elfHeader->name = part.name;
392 add(*part.elfHeader);
393
394 part.programHeaders =
395 std::make_unique<PartitionProgramHeadersSection<ELFT>>();
396 add(*part.programHeaders);
397 }
398
399 if (config->buildId != BuildIdKind::None) {
400 part.buildId = std::make_unique<BuildIdSection>();
401 add(*part.buildId);
402 }
403
404 part.dynStrTab = std::make_unique<StringTableSection>(".dynstr", true);
405 part.dynSymTab =
406 std::make_unique<SymbolTableSection<ELFT>>(*part.dynStrTab);
407 part.dynamic = std::make_unique<DynamicSection<ELFT>>();
408
409 if (hasMemtag()) {
410 part.memtagAndroidNote = std::make_unique<MemtagAndroidNote>();
411 add(*part.memtagAndroidNote);
412 if (canHaveMemtagGlobals()) {
413 part.memtagGlobalDescriptors =
414 std::make_unique<MemtagGlobalDescriptors>();
415 add(*part.memtagGlobalDescriptors);
416 }
417 }
418
419 if (config->androidPackDynRelocs)
420 part.relaDyn = std::make_unique<AndroidPackedRelocationSection<ELFT>>(
421 relaDynName, threadCount);
422 else
423 part.relaDyn = std::make_unique<RelocationSection<ELFT>>(
424 relaDynName, config->zCombreloc, threadCount);
425
426 if (config->hasDynSymTab) {
427 add(*part.dynSymTab);
428
429 part.verSym = std::make_unique<VersionTableSection>();
430 add(*part.verSym);
431
432 if (!namedVersionDefs().empty()) {
433 part.verDef = std::make_unique<VersionDefinitionSection>();
434 add(*part.verDef);
435 }
436
437 part.verNeed = std::make_unique<VersionNeedSection<ELFT>>();
438 add(*part.verNeed);
439
440 if (config->gnuHash) {
441 part.gnuHashTab = std::make_unique<GnuHashTableSection>();
442 add(*part.gnuHashTab);
443 }
444
445 if (config->sysvHash) {
446 part.hashTab = std::make_unique<HashTableSection>();
447 add(*part.hashTab);
448 }
449
450 add(*part.dynamic);
451 add(*part.dynStrTab);
452 add(*part.relaDyn);
453 }
454
455 if (config->relrPackDynRelocs) {
456 part.relrDyn = std::make_unique<RelrSection<ELFT>>(threadCount);
457 add(*part.relrDyn);
458 }
459
460 if (!config->relocatable) {
461 if (config->ehFrameHdr) {
462 part.ehFrameHdr = std::make_unique<EhFrameHeader>();
463 add(*part.ehFrameHdr);
464 }
465 part.ehFrame = std::make_unique<EhFrameSection>();
466 add(*part.ehFrame);
467
468 if (config->emachine == EM_ARM) {
469 // This section replaces all the individual .ARM.exidx InputSections.
470 part.armExidx = std::make_unique<ARMExidxSyntheticSection>();
471 add(*part.armExidx);
472 }
473 }
474
475 if (!config->packageMetadata.empty()) {
476 part.packageMetadataNote = std::make_unique<PackageMetadataNote>();
477 add(*part.packageMetadataNote);
478 }
479 }
480
481 if (partitions.size() != 1) {
482 // Create the partition end marker. This needs to be in partition number 255
483 // so that it is sorted after all other partitions. It also has other
484 // special handling (see createPhdrs() and combineEhSections()).
485 in.partEnd =
486 std::make_unique<BssSection>(".part.end", config->maxPageSize, 1);
487 in.partEnd->partition = 255;
488 add(*in.partEnd);
489
490 in.partIndex = std::make_unique<PartitionIndexSection>();
491 addOptionalRegular("__part_index_begin", in.partIndex.get(), 0);
492 addOptionalRegular("__part_index_end", in.partIndex.get(),
493 in.partIndex->getSize());
494 add(*in.partIndex);
495 }
496
497 // Add .got. MIPS' .got is so different from the other archs,
498 // it has its own class.
499 if (config->emachine == EM_MIPS) {
500 in.mipsGot = std::make_unique<MipsGotSection>();
501 add(*in.mipsGot);
502 } else {
503 in.got = std::make_unique<GotSection>();
504 add(*in.got);
505 }
506
507 if (config->emachine == EM_PPC) {
508 in.ppc32Got2 = std::make_unique<PPC32Got2Section>();
509 add(*in.ppc32Got2);
510 }
511
512 if (config->emachine == EM_PPC64) {
513 in.ppc64LongBranchTarget = std::make_unique<PPC64LongBranchTargetSection>();
514 add(*in.ppc64LongBranchTarget);
515 }
516
517 in.gotPlt = std::make_unique<GotPltSection>();
518 add(*in.gotPlt);
519 in.igotPlt = std::make_unique<IgotPltSection>();
520 add(*in.igotPlt);
521 // Add .relro_padding if DATA_SEGMENT_RELRO_END is used; otherwise, add the
522 // section in the absence of PHDRS/SECTIONS commands.
523 if (config->zRelro && ((script->phdrsCommands.empty() &&
524 !script->hasSectionsCommand) || script->seenRelroEnd)) {
525 in.relroPadding = std::make_unique<RelroPaddingSection>();
526 add(*in.relroPadding);
527 }
528
529 if (config->emachine == EM_ARM) {
530 in.armCmseSGSection = std::make_unique<ArmCmseSGSection>();
531 add(*in.armCmseSGSection);
532 }
533
534 // _GLOBAL_OFFSET_TABLE_ is defined relative to either .got.plt or .got. Treat
535 // it as a relocation and ensure the referenced section is created.
536 if (ElfSym::globalOffsetTable && config->emachine != EM_MIPS) {
537 if (target->gotBaseSymInGotPlt)
538 in.gotPlt->hasGotPltOffRel = true;
539 else
540 in.got->hasGotOffRel = true;
541 }
542
543 if (config->gdbIndex)
544 add(*GdbIndexSection::create<ELFT>());
545
546 // We always need to add rel[a].plt to output if it has entries.
547 // Even for static linking it can contain R_[*]_IRELATIVE relocations.
548 in.relaPlt = std::make_unique<RelocationSection<ELFT>>(
549 config->isRela ? ".rela.plt" : ".rel.plt", /*sort=*/false,
550 /*threadCount=*/1);
551 add(*in.relaPlt);
552
553 // The relaIplt immediately follows .rel[a].dyn to ensure that the IRelative
554 // relocations are processed last by the dynamic loader. We cannot place the
555 // iplt section in .rel.dyn when Android relocation packing is enabled because
556 // that would cause a section type mismatch. However, because the Android
557 // dynamic loader reads .rel.plt after .rel.dyn, we can get the desired
558 // behaviour by placing the iplt section in .rel.plt.
559 in.relaIplt = std::make_unique<RelocationSection<ELFT>>(
560 config->androidPackDynRelocs ? in.relaPlt->name : relaDynName,
561 /*sort=*/false, /*threadCount=*/1);
562 add(*in.relaIplt);
563
564 if ((config->emachine == EM_386 || config->emachine == EM_X86_64) &&
565 (config->andFeatures & GNU_PROPERTY_X86_FEATURE_1_IBT)) {
566 in.ibtPlt = std::make_unique<IBTPltSection>();
567 add(*in.ibtPlt);
568 }
569
570 if (config->emachine == EM_PPC)
571 in.plt = std::make_unique<PPC32GlinkSection>();
572 else
573 in.plt = std::make_unique<PltSection>();
574 add(*in.plt);
575 in.iplt = std::make_unique<IpltSection>();
576 add(*in.iplt);
577
578 if (config->andFeatures)
579 add(*make<GnuPropertySection>());
580
581 // .note.GNU-stack is always added when we are creating a re-linkable
582 // object file. Other linkers are using the presence of this marker
583 // section to control the executable-ness of the stack area, but that
584 // is irrelevant these days. Stack area should always be non-executable
585 // by default. So we emit this section unconditionally.
586 if (config->relocatable)
587 add(*make<GnuStackSection>());
588
589 if (in.symTab)
590 add(*in.symTab);
591 if (in.symTabShndx)
592 add(*in.symTabShndx);
593 add(*in.shStrTab);
594 if (in.strTab)
595 add(*in.strTab);
596 }
597
598 // The main function of the writer.
run()599 template <class ELFT> void Writer<ELFT>::run() {
600 // Now that we have a complete set of output sections. This function
601 // completes section contents. For example, we need to add strings
602 // to the string table, and add entries to .got and .plt.
603 // finalizeSections does that.
604 finalizeSections();
605 checkExecuteOnly();
606
607 // If --compressed-debug-sections is specified, compress .debug_* sections.
608 // Do it right now because it changes the size of output sections.
609 for (OutputSection *sec : outputSections)
610 sec->maybeCompress<ELFT>();
611
612 if (script->hasSectionsCommand)
613 script->allocateHeaders(mainPart->phdrs);
614
615 // Remove empty PT_LOAD to avoid causing the dynamic linker to try to mmap a
616 // 0 sized region. This has to be done late since only after assignAddresses
617 // we know the size of the sections.
618 for (Partition &part : partitions)
619 removeEmptyPTLoad(part.phdrs);
620
621 if (!config->oFormatBinary)
622 assignFileOffsets();
623 else
624 assignFileOffsetsBinary();
625
626 for (Partition &part : partitions)
627 setPhdrs(part);
628
629 // Handle --print-map(-M)/--Map and --cref. Dump them before checkSections()
630 // because the files may be useful in case checkSections() or openFile()
631 // fails, for example, due to an erroneous file size.
632 writeMapAndCref();
633
634 // Handle --print-memory-usage option.
635 if (config->printMemoryUsage)
636 script->printMemoryUsage(lld::outs());
637
638 if (config->checkSections)
639 checkSections();
640
641 // It does not make sense try to open the file if we have error already.
642 if (errorCount())
643 return;
644
645 {
646 llvm::TimeTraceScope timeScope("Write output file");
647 // Write the result down to a file.
648 openFile();
649 if (errorCount())
650 return;
651
652 if (!config->oFormatBinary) {
653 if (config->zSeparate != SeparateSegmentKind::None)
654 writeTrapInstr();
655 writeHeader();
656 writeSections();
657 } else {
658 writeSectionsBinary();
659 }
660
661 // Backfill .note.gnu.build-id section content. This is done at last
662 // because the content is usually a hash value of the entire output file.
663 writeBuildId();
664 if (errorCount())
665 return;
666
667 if (auto e = buffer->commit())
668 fatal("failed to write output '" + buffer->getPath() +
669 "': " + toString(std::move(e)));
670
671 if (!config->cmseOutputLib.empty())
672 writeARMCmseImportLib<ELFT>();
673 }
674 }
675
676 template <class ELFT, class RelTy>
markUsedLocalSymbolsImpl(ObjFile<ELFT> * file,llvm::ArrayRef<RelTy> rels)677 static void markUsedLocalSymbolsImpl(ObjFile<ELFT> *file,
678 llvm::ArrayRef<RelTy> rels) {
679 for (const RelTy &rel : rels) {
680 Symbol &sym = file->getRelocTargetSym(rel);
681 if (sym.isLocal())
682 sym.used = true;
683 }
684 }
685
686 // The function ensures that the "used" field of local symbols reflects the fact
687 // that the symbol is used in a relocation from a live section.
markUsedLocalSymbols()688 template <class ELFT> static void markUsedLocalSymbols() {
689 // With --gc-sections, the field is already filled.
690 // See MarkLive<ELFT>::resolveReloc().
691 if (config->gcSections)
692 return;
693 for (ELFFileBase *file : ctx.objectFiles) {
694 ObjFile<ELFT> *f = cast<ObjFile<ELFT>>(file);
695 for (InputSectionBase *s : f->getSections()) {
696 InputSection *isec = dyn_cast_or_null<InputSection>(s);
697 if (!isec)
698 continue;
699 if (isec->type == SHT_REL)
700 markUsedLocalSymbolsImpl(f, isec->getDataAs<typename ELFT::Rel>());
701 else if (isec->type == SHT_RELA)
702 markUsedLocalSymbolsImpl(f, isec->getDataAs<typename ELFT::Rela>());
703 }
704 }
705 }
706
shouldKeepInSymtab(const Defined & sym)707 static bool shouldKeepInSymtab(const Defined &sym) {
708 if (sym.isSection())
709 return false;
710
711 // If --emit-reloc or -r is given, preserve symbols referenced by relocations
712 // from live sections.
713 if (sym.used && config->copyRelocs)
714 return true;
715
716 // Exclude local symbols pointing to .ARM.exidx sections.
717 // They are probably mapping symbols "$d", which are optional for these
718 // sections. After merging the .ARM.exidx sections, some of these symbols
719 // may become dangling. The easiest way to avoid the issue is not to add
720 // them to the symbol table from the beginning.
721 if (config->emachine == EM_ARM && sym.section &&
722 sym.section->type == SHT_ARM_EXIDX)
723 return false;
724
725 if (config->discard == DiscardPolicy::None)
726 return true;
727 if (config->discard == DiscardPolicy::All)
728 return false;
729
730 // In ELF assembly .L symbols are normally discarded by the assembler.
731 // If the assembler fails to do so, the linker discards them if
732 // * --discard-locals is used.
733 // * The symbol is in a SHF_MERGE section, which is normally the reason for
734 // the assembler keeping the .L symbol.
735 if (sym.getName().starts_with(".L") &&
736 (config->discard == DiscardPolicy::Locals ||
737 (sym.section && (sym.section->flags & SHF_MERGE))))
738 return false;
739 return true;
740 }
741
includeInSymtab(const Symbol & b)742 bool lld::elf::includeInSymtab(const Symbol &b) {
743 if (auto *d = dyn_cast<Defined>(&b)) {
744 // Always include absolute symbols.
745 SectionBase *sec = d->section;
746 if (!sec)
747 return true;
748 assert(sec->isLive());
749
750 if (auto *s = dyn_cast<MergeInputSection>(sec))
751 return s->getSectionPiece(d->value).live;
752 return true;
753 }
754 return b.used || !config->gcSections;
755 }
756
757 // Scan local symbols to:
758 //
759 // - demote symbols defined relative to /DISCARD/ discarded input sections so
760 // that relocations referencing them will lead to errors.
761 // - copy eligible symbols to .symTab
demoteAndCopyLocalSymbols()762 static void demoteAndCopyLocalSymbols() {
763 llvm::TimeTraceScope timeScope("Add local symbols");
764 for (ELFFileBase *file : ctx.objectFiles) {
765 DenseMap<SectionBase *, size_t> sectionIndexMap;
766 for (Symbol *b : file->getLocalSymbols()) {
767 assert(b->isLocal() && "should have been caught in initializeSymbols()");
768 auto *dr = dyn_cast<Defined>(b);
769 if (!dr)
770 continue;
771
772 if (dr->section && !dr->section->isLive())
773 demoteDefined(*dr, sectionIndexMap);
774 else if (in.symTab && includeInSymtab(*b) && shouldKeepInSymtab(*dr))
775 in.symTab->addSymbol(b);
776 }
777 }
778 }
779
780 // Create a section symbol for each output section so that we can represent
781 // relocations that point to the section. If we know that no relocation is
782 // referring to a section (that happens if the section is a synthetic one), we
783 // don't create a section symbol for that section.
addSectionSymbols()784 template <class ELFT> void Writer<ELFT>::addSectionSymbols() {
785 for (SectionCommand *cmd : script->sectionCommands) {
786 auto *osd = dyn_cast<OutputDesc>(cmd);
787 if (!osd)
788 continue;
789 OutputSection &osec = osd->osec;
790 InputSectionBase *isec = nullptr;
791 // Iterate over all input sections and add a STT_SECTION symbol if any input
792 // section may be a relocation target.
793 for (SectionCommand *cmd : osec.commands) {
794 auto *isd = dyn_cast<InputSectionDescription>(cmd);
795 if (!isd)
796 continue;
797 for (InputSectionBase *s : isd->sections) {
798 // Relocations are not using REL[A] section symbols.
799 if (s->type == SHT_REL || s->type == SHT_RELA)
800 continue;
801
802 // Unlike other synthetic sections, mergeable output sections contain
803 // data copied from input sections, and there may be a relocation
804 // pointing to its contents if -r or --emit-reloc is given.
805 if (isa<SyntheticSection>(s) && !(s->flags & SHF_MERGE))
806 continue;
807
808 isec = s;
809 break;
810 }
811 }
812 if (!isec)
813 continue;
814
815 // Set the symbol to be relative to the output section so that its st_value
816 // equals the output section address. Note, there may be a gap between the
817 // start of the output section and isec.
818 in.symTab->addSymbol(makeDefined(isec->file, "", STB_LOCAL, /*stOther=*/0,
819 STT_SECTION,
820 /*value=*/0, /*size=*/0, &osec));
821 }
822 }
823
824 // Today's loaders have a feature to make segments read-only after
825 // processing dynamic relocations to enhance security. PT_GNU_RELRO
826 // is defined for that.
827 //
828 // This function returns true if a section needs to be put into a
829 // PT_GNU_RELRO segment.
isRelroSection(const OutputSection * sec)830 static bool isRelroSection(const OutputSection *sec) {
831 if (!config->zRelro)
832 return false;
833 if (sec->relro)
834 return true;
835
836 uint64_t flags = sec->flags;
837
838 // Non-allocatable or non-writable sections don't need RELRO because
839 // they are not writable or not even mapped to memory in the first place.
840 // RELRO is for sections that are essentially read-only but need to
841 // be writable only at process startup to allow dynamic linker to
842 // apply relocations.
843 if (!(flags & SHF_ALLOC) || !(flags & SHF_WRITE))
844 return false;
845
846 // Once initialized, TLS data segments are used as data templates
847 // for a thread-local storage. For each new thread, runtime
848 // allocates memory for a TLS and copy templates there. No thread
849 // are supposed to use templates directly. Thus, it can be in RELRO.
850 if (flags & SHF_TLS)
851 return true;
852
853 // .init_array, .preinit_array and .fini_array contain pointers to
854 // functions that are executed on process startup or exit. These
855 // pointers are set by the static linker, and they are not expected
856 // to change at runtime. But if you are an attacker, you could do
857 // interesting things by manipulating pointers in .fini_array, for
858 // example. So they are put into RELRO.
859 uint32_t type = sec->type;
860 if (type == SHT_INIT_ARRAY || type == SHT_FINI_ARRAY ||
861 type == SHT_PREINIT_ARRAY)
862 return true;
863
864 // .got contains pointers to external symbols. They are resolved by
865 // the dynamic linker when a module is loaded into memory, and after
866 // that they are not expected to change. So, it can be in RELRO.
867 if (in.got && sec == in.got->getParent())
868 return true;
869
870 // .toc is a GOT-ish section for PowerPC64. Their contents are accessed
871 // through r2 register, which is reserved for that purpose. Since r2 is used
872 // for accessing .got as well, .got and .toc need to be close enough in the
873 // virtual address space. Usually, .toc comes just after .got. Since we place
874 // .got into RELRO, .toc needs to be placed into RELRO too.
875 if (sec->name.equals(".toc"))
876 return true;
877
878 // .got.plt contains pointers to external function symbols. They are
879 // by default resolved lazily, so we usually cannot put it into RELRO.
880 // However, if "-z now" is given, the lazy symbol resolution is
881 // disabled, which enables us to put it into RELRO.
882 if (sec == in.gotPlt->getParent())
883 return config->zNow;
884
885 if (in.relroPadding && sec == in.relroPadding->getParent())
886 return true;
887
888 // .dynamic section contains data for the dynamic linker, and
889 // there's no need to write to it at runtime, so it's better to put
890 // it into RELRO.
891 if (sec->name == ".dynamic")
892 return true;
893
894 // Sections with some special names are put into RELRO. This is a
895 // bit unfortunate because section names shouldn't be significant in
896 // ELF in spirit. But in reality many linker features depend on
897 // magic section names.
898 StringRef s = sec->name;
899 return s == ".data.rel.ro" || s == ".bss.rel.ro" || s == ".ctors" ||
900 s == ".dtors" || s == ".jcr" || s == ".eh_frame" ||
901 s == ".fini_array" || s == ".init_array" ||
902 s == ".openbsd.randomdata" || s == ".preinit_array";
903 }
904
905 // We compute a rank for each section. The rank indicates where the
906 // section should be placed in the file. Instead of using simple
907 // numbers (0,1,2...), we use a series of flags. One for each decision
908 // point when placing the section.
909 // Using flags has two key properties:
910 // * It is easy to check if a give branch was taken.
911 // * It is easy two see how similar two ranks are (see getRankProximity).
912 enum RankFlags {
913 RF_NOT_ADDR_SET = 1 << 27,
914 RF_NOT_ALLOC = 1 << 26,
915 RF_PARTITION = 1 << 18, // Partition number (8 bits)
916 RF_NOT_SPECIAL = 1 << 17,
917 RF_WRITE = 1 << 16,
918 RF_EXEC_WRITE = 1 << 15,
919 RF_EXEC = 1 << 14,
920 RF_RODATA = 1 << 13,
921 RF_LARGE = 1 << 12,
922 RF_NOT_RELRO = 1 << 9,
923 RF_NOT_TLS = 1 << 8,
924 RF_BSS = 1 << 7,
925 };
926
getSectionRank(OutputSection & osec)927 static unsigned getSectionRank(OutputSection &osec) {
928 unsigned rank = osec.partition * RF_PARTITION;
929
930 // We want to put section specified by -T option first, so we
931 // can start assigning VA starting from them later.
932 if (config->sectionStartMap.count(osec.name))
933 return rank;
934 rank |= RF_NOT_ADDR_SET;
935
936 // Allocatable sections go first to reduce the total PT_LOAD size and
937 // so debug info doesn't change addresses in actual code.
938 if (!(osec.flags & SHF_ALLOC))
939 return rank | RF_NOT_ALLOC;
940
941 if (osec.type == SHT_LLVM_PART_EHDR)
942 return rank;
943 if (osec.type == SHT_LLVM_PART_PHDR)
944 return rank | 1;
945
946 // Put .interp first because some loaders want to see that section
947 // on the first page of the executable file when loaded into memory.
948 if (osec.name == ".interp")
949 return rank | 2;
950
951 // Put .note sections at the beginning so that they are likely to be included
952 // in a truncate core file. In particular, .note.gnu.build-id, if available,
953 // can identify the object file.
954 if (osec.type == SHT_NOTE)
955 return rank | 3;
956
957 rank |= RF_NOT_SPECIAL;
958
959 // Sort sections based on their access permission in the following
960 // order: R, RX, RXW, RW(RELRO), RW(non-RELRO).
961 //
962 // Read-only sections come first such that they go in the PT_LOAD covering the
963 // program headers at the start of the file.
964 //
965 // The layout for writable sections is PT_LOAD(PT_GNU_RELRO(.data.rel.ro
966 // .bss.rel.ro) | .data .bss), where | marks where page alignment happens.
967 // An alternative ordering is PT_LOAD(.data | PT_GNU_RELRO( .data.rel.ro
968 // .bss.rel.ro) | .bss), but it may waste more bytes due to 2 alignment
969 // places.
970 bool isExec = osec.flags & SHF_EXECINSTR;
971 bool isWrite = osec.flags & SHF_WRITE;
972
973 if (!isWrite && !isExec) {
974 // Make PROGBITS sections (e.g .rodata .eh_frame) closer to .text to
975 // alleviate relocation overflow pressure. Large special sections such as
976 // .dynstr and .dynsym can be away from .text.
977 if (osec.type == SHT_PROGBITS)
978 rank |= RF_RODATA;
979 // Among PROGBITS sections, place .lrodata further from .text.
980 if (!(osec.flags & SHF_X86_64_LARGE && config->emachine == EM_X86_64))
981 rank |= RF_LARGE;
982 } else if (isExec) {
983 rank |= isWrite ? RF_EXEC_WRITE : RF_EXEC;
984 } else {
985 rank |= RF_WRITE;
986 // The TLS initialization block needs to be a single contiguous block. Place
987 // TLS sections directly before the other RELRO sections.
988 if (!(osec.flags & SHF_TLS))
989 rank |= RF_NOT_TLS;
990 if (isRelroSection(&osec))
991 osec.relro = true;
992 else
993 rank |= RF_NOT_RELRO;
994 // Place .ldata and .lbss after .bss. Making .bss closer to .text alleviates
995 // relocation overflow pressure.
996 if (osec.flags & SHF_X86_64_LARGE && config->emachine == EM_X86_64)
997 rank |= RF_LARGE;
998 }
999
1000 // Within TLS sections, or within other RelRo sections, or within non-RelRo
1001 // sections, place non-NOBITS sections first.
1002 if (osec.type == SHT_NOBITS)
1003 rank |= RF_BSS;
1004
1005 // Some architectures have additional ordering restrictions for sections
1006 // within the same PT_LOAD.
1007 if (config->emachine == EM_PPC64) {
1008 // PPC64 has a number of special SHT_PROGBITS+SHF_ALLOC+SHF_WRITE sections
1009 // that we would like to make sure appear is a specific order to maximize
1010 // their coverage by a single signed 16-bit offset from the TOC base
1011 // pointer.
1012 StringRef name = osec.name;
1013 if (name == ".got")
1014 rank |= 1;
1015 else if (name == ".toc")
1016 rank |= 2;
1017 }
1018
1019 if (config->emachine == EM_MIPS) {
1020 if (osec.name != ".got")
1021 rank |= 1;
1022 // All sections with SHF_MIPS_GPREL flag should be grouped together
1023 // because data in these sections is addressable with a gp relative address.
1024 if (osec.flags & SHF_MIPS_GPREL)
1025 rank |= 2;
1026 }
1027
1028 if (config->emachine == EM_RISCV) {
1029 // .sdata and .sbss are placed closer to make GP relaxation more profitable
1030 // and match GNU ld.
1031 StringRef name = osec.name;
1032 if (name == ".sdata" || (osec.type == SHT_NOBITS && name != ".sbss"))
1033 rank |= 1;
1034 }
1035
1036 return rank;
1037 }
1038
compareSections(const SectionCommand * aCmd,const SectionCommand * bCmd)1039 static bool compareSections(const SectionCommand *aCmd,
1040 const SectionCommand *bCmd) {
1041 const OutputSection *a = &cast<OutputDesc>(aCmd)->osec;
1042 const OutputSection *b = &cast<OutputDesc>(bCmd)->osec;
1043
1044 if (a->sortRank != b->sortRank)
1045 return a->sortRank < b->sortRank;
1046
1047 if (!(a->sortRank & RF_NOT_ADDR_SET))
1048 return config->sectionStartMap.lookup(a->name) <
1049 config->sectionStartMap.lookup(b->name);
1050 return false;
1051 }
1052
add(OutputSection * sec)1053 void PhdrEntry::add(OutputSection *sec) {
1054 lastSec = sec;
1055 if (!firstSec)
1056 firstSec = sec;
1057 p_align = std::max(p_align, sec->addralign);
1058 if (p_type == PT_LOAD)
1059 sec->ptLoad = this;
1060 }
1061
1062 // The beginning and the ending of .rel[a].plt section are marked
1063 // with __rel[a]_iplt_{start,end} symbols if it is a statically linked
1064 // executable. The runtime needs these symbols in order to resolve
1065 // all IRELATIVE relocs on startup. For dynamic executables, we don't
1066 // need these symbols, since IRELATIVE relocs are resolved through GOT
1067 // and PLT. For details, see http://www.airs.com/blog/archives/403.
addRelIpltSymbols()1068 template <class ELFT> void Writer<ELFT>::addRelIpltSymbols() {
1069 if (config->isPic)
1070 return;
1071
1072 // By default, __rela_iplt_{start,end} belong to a dummy section 0
1073 // because .rela.plt might be empty and thus removed from output.
1074 // We'll override Out::elfHeader with In.relaIplt later when we are
1075 // sure that .rela.plt exists in output.
1076 ElfSym::relaIpltStart = addOptionalRegular(
1077 config->isRela ? "__rela_iplt_start" : "__rel_iplt_start",
1078 Out::elfHeader, 0, STV_HIDDEN);
1079
1080 ElfSym::relaIpltEnd = addOptionalRegular(
1081 config->isRela ? "__rela_iplt_end" : "__rel_iplt_end",
1082 Out::elfHeader, 0, STV_HIDDEN);
1083 }
1084
1085 // This function generates assignments for predefined symbols (e.g. _end or
1086 // _etext) and inserts them into the commands sequence to be processed at the
1087 // appropriate time. This ensures that the value is going to be correct by the
1088 // time any references to these symbols are processed and is equivalent to
1089 // defining these symbols explicitly in the linker script.
setReservedSymbolSections()1090 template <class ELFT> void Writer<ELFT>::setReservedSymbolSections() {
1091 if (ElfSym::globalOffsetTable) {
1092 // The _GLOBAL_OFFSET_TABLE_ symbol is defined by target convention usually
1093 // to the start of the .got or .got.plt section.
1094 InputSection *sec = in.gotPlt.get();
1095 if (!target->gotBaseSymInGotPlt)
1096 sec = in.mipsGot ? cast<InputSection>(in.mipsGot.get())
1097 : cast<InputSection>(in.got.get());
1098 ElfSym::globalOffsetTable->section = sec;
1099 }
1100
1101 // .rela_iplt_{start,end} mark the start and the end of in.relaIplt.
1102 if (ElfSym::relaIpltStart && in.relaIplt->isNeeded()) {
1103 ElfSym::relaIpltStart->section = in.relaIplt.get();
1104 ElfSym::relaIpltEnd->section = in.relaIplt.get();
1105 ElfSym::relaIpltEnd->value = in.relaIplt->getSize();
1106 }
1107
1108 PhdrEntry *last = nullptr;
1109 PhdrEntry *lastRO = nullptr;
1110
1111 for (Partition &part : partitions) {
1112 for (PhdrEntry *p : part.phdrs) {
1113 if (p->p_type != PT_LOAD)
1114 continue;
1115 last = p;
1116 if (!(p->p_flags & PF_W))
1117 lastRO = p;
1118 }
1119 }
1120
1121 if (lastRO) {
1122 // _etext is the first location after the last read-only loadable segment.
1123 if (ElfSym::etext1)
1124 ElfSym::etext1->section = lastRO->lastSec;
1125 if (ElfSym::etext2)
1126 ElfSym::etext2->section = lastRO->lastSec;
1127 }
1128
1129 if (last) {
1130 // _edata points to the end of the last mapped initialized section.
1131 OutputSection *edata = nullptr;
1132 for (OutputSection *os : outputSections) {
1133 if (os->type != SHT_NOBITS)
1134 edata = os;
1135 if (os == last->lastSec)
1136 break;
1137 }
1138
1139 if (ElfSym::edata1)
1140 ElfSym::edata1->section = edata;
1141 if (ElfSym::edata2)
1142 ElfSym::edata2->section = edata;
1143
1144 // _end is the first location after the uninitialized data region.
1145 if (ElfSym::end1)
1146 ElfSym::end1->section = last->lastSec;
1147 if (ElfSym::end2)
1148 ElfSym::end2->section = last->lastSec;
1149 }
1150
1151 if (ElfSym::bss) {
1152 // On RISC-V, set __bss_start to the start of .sbss if present.
1153 OutputSection *sbss =
1154 config->emachine == EM_RISCV ? findSection(".sbss") : nullptr;
1155 ElfSym::bss->section = sbss ? sbss : findSection(".bss");
1156 }
1157
1158 // Setup MIPS _gp_disp/__gnu_local_gp symbols which should
1159 // be equal to the _gp symbol's value.
1160 if (ElfSym::mipsGp) {
1161 // Find GP-relative section with the lowest address
1162 // and use this address to calculate default _gp value.
1163 for (OutputSection *os : outputSections) {
1164 if (os->flags & SHF_MIPS_GPREL) {
1165 ElfSym::mipsGp->section = os;
1166 ElfSym::mipsGp->value = 0x7ff0;
1167 break;
1168 }
1169 }
1170 }
1171 }
1172
1173 // We want to find how similar two ranks are.
1174 // The more branches in getSectionRank that match, the more similar they are.
1175 // Since each branch corresponds to a bit flag, we can just use
1176 // countLeadingZeros.
getRankProximity(OutputSection * a,SectionCommand * b)1177 static int getRankProximity(OutputSection *a, SectionCommand *b) {
1178 auto *osd = dyn_cast<OutputDesc>(b);
1179 return (osd && osd->osec.hasInputSections)
1180 ? llvm::countl_zero(a->sortRank ^ osd->osec.sortRank)
1181 : -1;
1182 }
1183
1184 // When placing orphan sections, we want to place them after symbol assignments
1185 // so that an orphan after
1186 // begin_foo = .;
1187 // foo : { *(foo) }
1188 // end_foo = .;
1189 // doesn't break the intended meaning of the begin/end symbols.
1190 // We don't want to go over sections since findOrphanPos is the
1191 // one in charge of deciding the order of the sections.
1192 // We don't want to go over changes to '.', since doing so in
1193 // rx_sec : { *(rx_sec) }
1194 // . = ALIGN(0x1000);
1195 // /* The RW PT_LOAD starts here*/
1196 // rw_sec : { *(rw_sec) }
1197 // would mean that the RW PT_LOAD would become unaligned.
shouldSkip(SectionCommand * cmd)1198 static bool shouldSkip(SectionCommand *cmd) {
1199 if (auto *assign = dyn_cast<SymbolAssignment>(cmd))
1200 return assign->name != ".";
1201 return false;
1202 }
1203
1204 // We want to place orphan sections so that they share as much
1205 // characteristics with their neighbors as possible. For example, if
1206 // both are rw, or both are tls.
1207 static SmallVectorImpl<SectionCommand *>::iterator
findOrphanPos(SmallVectorImpl<SectionCommand * >::iterator b,SmallVectorImpl<SectionCommand * >::iterator e)1208 findOrphanPos(SmallVectorImpl<SectionCommand *>::iterator b,
1209 SmallVectorImpl<SectionCommand *>::iterator e) {
1210 OutputSection *sec = &cast<OutputDesc>(*e)->osec;
1211
1212 // As a special case, place .relro_padding before the SymbolAssignment using
1213 // DATA_SEGMENT_RELRO_END, if present.
1214 if (in.relroPadding && sec == in.relroPadding->getParent()) {
1215 auto i = std::find_if(b, e, [=](SectionCommand *a) {
1216 if (auto *assign = dyn_cast<SymbolAssignment>(a))
1217 return assign->dataSegmentRelroEnd;
1218 return false;
1219 });
1220 if (i != e)
1221 return i;
1222 }
1223
1224 // Find the first element that has as close a rank as possible.
1225 auto i = std::max_element(b, e, [=](SectionCommand *a, SectionCommand *b) {
1226 return getRankProximity(sec, a) < getRankProximity(sec, b);
1227 });
1228 if (i == e)
1229 return e;
1230 if (!isa<OutputDesc>(*i))
1231 return e;
1232 auto foundSec = &cast<OutputDesc>(*i)->osec;
1233
1234 // Consider all existing sections with the same proximity.
1235 int proximity = getRankProximity(sec, *i);
1236 unsigned sortRank = sec->sortRank;
1237 if (script->hasPhdrsCommands() || !script->memoryRegions.empty())
1238 // Prevent the orphan section to be placed before the found section. If
1239 // custom program headers are defined, that helps to avoid adding it to a
1240 // previous segment and changing flags of that segment, for example, making
1241 // a read-only segment writable. If memory regions are defined, an orphan
1242 // section should continue the same region as the found section to better
1243 // resemble the behavior of GNU ld.
1244 sortRank = std::max(sortRank, foundSec->sortRank);
1245 for (; i != e; ++i) {
1246 auto *curSecDesc = dyn_cast<OutputDesc>(*i);
1247 if (!curSecDesc || !curSecDesc->osec.hasInputSections)
1248 continue;
1249 if (getRankProximity(sec, curSecDesc) != proximity ||
1250 sortRank < curSecDesc->osec.sortRank)
1251 break;
1252 }
1253
1254 auto isOutputSecWithInputSections = [](SectionCommand *cmd) {
1255 auto *osd = dyn_cast<OutputDesc>(cmd);
1256 return osd && osd->osec.hasInputSections;
1257 };
1258 auto j =
1259 std::find_if(std::make_reverse_iterator(i), std::make_reverse_iterator(b),
1260 isOutputSecWithInputSections);
1261 i = j.base();
1262
1263 // As a special case, if the orphan section is the last section, put
1264 // it at the very end, past any other commands.
1265 // This matches bfd's behavior and is convenient when the linker script fully
1266 // specifies the start of the file, but doesn't care about the end (the non
1267 // alloc sections for example).
1268 auto nextSec = std::find_if(i, e, isOutputSecWithInputSections);
1269 if (nextSec == e)
1270 return e;
1271
1272 while (i != e && shouldSkip(*i))
1273 ++i;
1274 return i;
1275 }
1276
1277 // Adds random priorities to sections not already in the map.
maybeShuffle(DenseMap<const InputSectionBase *,int> & order)1278 static void maybeShuffle(DenseMap<const InputSectionBase *, int> &order) {
1279 if (config->shuffleSections.empty())
1280 return;
1281
1282 SmallVector<InputSectionBase *, 0> matched, sections = ctx.inputSections;
1283 matched.reserve(sections.size());
1284 for (const auto &patAndSeed : config->shuffleSections) {
1285 matched.clear();
1286 for (InputSectionBase *sec : sections)
1287 if (patAndSeed.first.match(sec->name))
1288 matched.push_back(sec);
1289 const uint32_t seed = patAndSeed.second;
1290 if (seed == UINT32_MAX) {
1291 // If --shuffle-sections <section-glob>=-1, reverse the section order. The
1292 // section order is stable even if the number of sections changes. This is
1293 // useful to catch issues like static initialization order fiasco
1294 // reliably.
1295 std::reverse(matched.begin(), matched.end());
1296 } else {
1297 std::mt19937 g(seed ? seed : std::random_device()());
1298 llvm::shuffle(matched.begin(), matched.end(), g);
1299 }
1300 size_t i = 0;
1301 for (InputSectionBase *&sec : sections)
1302 if (patAndSeed.first.match(sec->name))
1303 sec = matched[i++];
1304 }
1305
1306 // Existing priorities are < 0, so use priorities >= 0 for the missing
1307 // sections.
1308 int prio = 0;
1309 for (InputSectionBase *sec : sections) {
1310 if (order.try_emplace(sec, prio).second)
1311 ++prio;
1312 }
1313 }
1314
1315 // Builds section order for handling --symbol-ordering-file.
buildSectionOrder()1316 static DenseMap<const InputSectionBase *, int> buildSectionOrder() {
1317 DenseMap<const InputSectionBase *, int> sectionOrder;
1318 // Use the rarely used option --call-graph-ordering-file to sort sections.
1319 if (!config->callGraphProfile.empty())
1320 return computeCallGraphProfileOrder();
1321
1322 if (config->symbolOrderingFile.empty())
1323 return sectionOrder;
1324
1325 struct SymbolOrderEntry {
1326 int priority;
1327 bool present;
1328 };
1329
1330 // Build a map from symbols to their priorities. Symbols that didn't
1331 // appear in the symbol ordering file have the lowest priority 0.
1332 // All explicitly mentioned symbols have negative (higher) priorities.
1333 DenseMap<CachedHashStringRef, SymbolOrderEntry> symbolOrder;
1334 int priority = -config->symbolOrderingFile.size();
1335 for (StringRef s : config->symbolOrderingFile)
1336 symbolOrder.insert({CachedHashStringRef(s), {priority++, false}});
1337
1338 // Build a map from sections to their priorities.
1339 auto addSym = [&](Symbol &sym) {
1340 auto it = symbolOrder.find(CachedHashStringRef(sym.getName()));
1341 if (it == symbolOrder.end())
1342 return;
1343 SymbolOrderEntry &ent = it->second;
1344 ent.present = true;
1345
1346 maybeWarnUnorderableSymbol(&sym);
1347
1348 if (auto *d = dyn_cast<Defined>(&sym)) {
1349 if (auto *sec = dyn_cast_or_null<InputSectionBase>(d->section)) {
1350 int &priority = sectionOrder[cast<InputSectionBase>(sec)];
1351 priority = std::min(priority, ent.priority);
1352 }
1353 }
1354 };
1355
1356 // We want both global and local symbols. We get the global ones from the
1357 // symbol table and iterate the object files for the local ones.
1358 for (Symbol *sym : symtab.getSymbols())
1359 addSym(*sym);
1360
1361 for (ELFFileBase *file : ctx.objectFiles)
1362 for (Symbol *sym : file->getLocalSymbols())
1363 addSym(*sym);
1364
1365 if (config->warnSymbolOrdering)
1366 for (auto orderEntry : symbolOrder)
1367 if (!orderEntry.second.present)
1368 warn("symbol ordering file: no such symbol: " + orderEntry.first.val());
1369
1370 return sectionOrder;
1371 }
1372
1373 // Sorts the sections in ISD according to the provided section order.
1374 static void
sortISDBySectionOrder(InputSectionDescription * isd,const DenseMap<const InputSectionBase *,int> & order,bool executableOutputSection)1375 sortISDBySectionOrder(InputSectionDescription *isd,
1376 const DenseMap<const InputSectionBase *, int> &order,
1377 bool executableOutputSection) {
1378 SmallVector<InputSection *, 0> unorderedSections;
1379 SmallVector<std::pair<InputSection *, int>, 0> orderedSections;
1380 uint64_t unorderedSize = 0;
1381 uint64_t totalSize = 0;
1382
1383 for (InputSection *isec : isd->sections) {
1384 if (executableOutputSection)
1385 totalSize += isec->getSize();
1386 auto i = order.find(isec);
1387 if (i == order.end()) {
1388 unorderedSections.push_back(isec);
1389 unorderedSize += isec->getSize();
1390 continue;
1391 }
1392 orderedSections.push_back({isec, i->second});
1393 }
1394 llvm::sort(orderedSections, llvm::less_second());
1395
1396 // Find an insertion point for the ordered section list in the unordered
1397 // section list. On targets with limited-range branches, this is the mid-point
1398 // of the unordered section list. This decreases the likelihood that a range
1399 // extension thunk will be needed to enter or exit the ordered region. If the
1400 // ordered section list is a list of hot functions, we can generally expect
1401 // the ordered functions to be called more often than the unordered functions,
1402 // making it more likely that any particular call will be within range, and
1403 // therefore reducing the number of thunks required.
1404 //
1405 // For example, imagine that you have 8MB of hot code and 32MB of cold code.
1406 // If the layout is:
1407 //
1408 // 8MB hot
1409 // 32MB cold
1410 //
1411 // only the first 8-16MB of the cold code (depending on which hot function it
1412 // is actually calling) can call the hot code without a range extension thunk.
1413 // However, if we use this layout:
1414 //
1415 // 16MB cold
1416 // 8MB hot
1417 // 16MB cold
1418 //
1419 // both the last 8-16MB of the first block of cold code and the first 8-16MB
1420 // of the second block of cold code can call the hot code without a thunk. So
1421 // we effectively double the amount of code that could potentially call into
1422 // the hot code without a thunk.
1423 //
1424 // The above is not necessary if total size of input sections in this "isd"
1425 // is small. Note that we assume all input sections are executable if the
1426 // output section is executable (which is not always true but supposed to
1427 // cover most cases).
1428 size_t insPt = 0;
1429 if (executableOutputSection && !orderedSections.empty() &&
1430 target->getThunkSectionSpacing() &&
1431 totalSize >= target->getThunkSectionSpacing()) {
1432 uint64_t unorderedPos = 0;
1433 for (; insPt != unorderedSections.size(); ++insPt) {
1434 unorderedPos += unorderedSections[insPt]->getSize();
1435 if (unorderedPos > unorderedSize / 2)
1436 break;
1437 }
1438 }
1439
1440 isd->sections.clear();
1441 for (InputSection *isec : ArrayRef(unorderedSections).slice(0, insPt))
1442 isd->sections.push_back(isec);
1443 for (std::pair<InputSection *, int> p : orderedSections)
1444 isd->sections.push_back(p.first);
1445 for (InputSection *isec : ArrayRef(unorderedSections).slice(insPt))
1446 isd->sections.push_back(isec);
1447 }
1448
sortSection(OutputSection & osec,const DenseMap<const InputSectionBase *,int> & order)1449 static void sortSection(OutputSection &osec,
1450 const DenseMap<const InputSectionBase *, int> &order) {
1451 StringRef name = osec.name;
1452
1453 // Never sort these.
1454 if (name == ".init" || name == ".fini")
1455 return;
1456
1457 // IRelative relocations that usually live in the .rel[a].dyn section should
1458 // be processed last by the dynamic loader. To achieve that we add synthetic
1459 // sections in the required order from the beginning so that the in.relaIplt
1460 // section is placed last in an output section. Here we just do not apply
1461 // sorting for an output section which holds the in.relaIplt section.
1462 if (in.relaIplt->getParent() == &osec)
1463 return;
1464
1465 // Sort input sections by priority using the list provided by
1466 // --symbol-ordering-file or --shuffle-sections=. This is a least significant
1467 // digit radix sort. The sections may be sorted stably again by a more
1468 // significant key.
1469 if (!order.empty())
1470 for (SectionCommand *b : osec.commands)
1471 if (auto *isd = dyn_cast<InputSectionDescription>(b))
1472 sortISDBySectionOrder(isd, order, osec.flags & SHF_EXECINSTR);
1473
1474 if (script->hasSectionsCommand)
1475 return;
1476
1477 if (name == ".init_array" || name == ".fini_array") {
1478 osec.sortInitFini();
1479 } else if (name == ".ctors" || name == ".dtors") {
1480 osec.sortCtorsDtors();
1481 } else if (config->emachine == EM_PPC64 && name == ".toc") {
1482 // .toc is allocated just after .got and is accessed using GOT-relative
1483 // relocations. Object files compiled with small code model have an
1484 // addressable range of [.got, .got + 0xFFFC] for GOT-relative relocations.
1485 // To reduce the risk of relocation overflow, .toc contents are sorted so
1486 // that sections having smaller relocation offsets are at beginning of .toc
1487 assert(osec.commands.size() == 1);
1488 auto *isd = cast<InputSectionDescription>(osec.commands[0]);
1489 llvm::stable_sort(isd->sections,
1490 [](const InputSection *a, const InputSection *b) -> bool {
1491 return a->file->ppc64SmallCodeModelTocRelocs &&
1492 !b->file->ppc64SmallCodeModelTocRelocs;
1493 });
1494 }
1495 }
1496
1497 // If no layout was provided by linker script, we want to apply default
1498 // sorting for special input sections. This also handles --symbol-ordering-file.
sortInputSections()1499 template <class ELFT> void Writer<ELFT>::sortInputSections() {
1500 // Build the order once since it is expensive.
1501 DenseMap<const InputSectionBase *, int> order = buildSectionOrder();
1502 maybeShuffle(order);
1503 for (SectionCommand *cmd : script->sectionCommands)
1504 if (auto *osd = dyn_cast<OutputDesc>(cmd))
1505 sortSection(osd->osec, order);
1506 }
1507
sortSections()1508 template <class ELFT> void Writer<ELFT>::sortSections() {
1509 llvm::TimeTraceScope timeScope("Sort sections");
1510
1511 // Don't sort if using -r. It is not necessary and we want to preserve the
1512 // relative order for SHF_LINK_ORDER sections.
1513 if (config->relocatable) {
1514 script->adjustOutputSections();
1515 return;
1516 }
1517
1518 sortInputSections();
1519
1520 for (SectionCommand *cmd : script->sectionCommands)
1521 if (auto *osd = dyn_cast<OutputDesc>(cmd))
1522 osd->osec.sortRank = getSectionRank(osd->osec);
1523 if (!script->hasSectionsCommand) {
1524 // OutputDescs are mostly contiguous, but may be interleaved with
1525 // SymbolAssignments in the presence of INSERT commands.
1526 auto mid = std::stable_partition(
1527 script->sectionCommands.begin(), script->sectionCommands.end(),
1528 [](SectionCommand *cmd) { return isa<OutputDesc>(cmd); });
1529 std::stable_sort(script->sectionCommands.begin(), mid, compareSections);
1530 }
1531
1532 // Process INSERT commands and update output section attributes. From this
1533 // point onwards the order of script->sectionCommands is fixed.
1534 script->processInsertCommands();
1535 script->adjustOutputSections();
1536
1537 if (script->hasSectionsCommand)
1538 sortOrphanSections();
1539
1540 script->adjustSectionsAfterSorting();
1541 }
1542
sortOrphanSections()1543 template <class ELFT> void Writer<ELFT>::sortOrphanSections() {
1544 // Orphan sections are sections present in the input files which are
1545 // not explicitly placed into the output file by the linker script.
1546 //
1547 // The sections in the linker script are already in the correct
1548 // order. We have to figuere out where to insert the orphan
1549 // sections.
1550 //
1551 // The order of the sections in the script is arbitrary and may not agree with
1552 // compareSections. This means that we cannot easily define a strict weak
1553 // ordering. To see why, consider a comparison of a section in the script and
1554 // one not in the script. We have a two simple options:
1555 // * Make them equivalent (a is not less than b, and b is not less than a).
1556 // The problem is then that equivalence has to be transitive and we can
1557 // have sections a, b and c with only b in a script and a less than c
1558 // which breaks this property.
1559 // * Use compareSectionsNonScript. Given that the script order doesn't have
1560 // to match, we can end up with sections a, b, c, d where b and c are in the
1561 // script and c is compareSectionsNonScript less than b. In which case d
1562 // can be equivalent to c, a to b and d < a. As a concrete example:
1563 // .a (rx) # not in script
1564 // .b (rx) # in script
1565 // .c (ro) # in script
1566 // .d (ro) # not in script
1567 //
1568 // The way we define an order then is:
1569 // * Sort only the orphan sections. They are in the end right now.
1570 // * Move each orphan section to its preferred position. We try
1571 // to put each section in the last position where it can share
1572 // a PT_LOAD.
1573 //
1574 // There is some ambiguity as to where exactly a new entry should be
1575 // inserted, because Commands contains not only output section
1576 // commands but also other types of commands such as symbol assignment
1577 // expressions. There's no correct answer here due to the lack of the
1578 // formal specification of the linker script. We use heuristics to
1579 // determine whether a new output command should be added before or
1580 // after another commands. For the details, look at shouldSkip
1581 // function.
1582
1583 auto i = script->sectionCommands.begin();
1584 auto e = script->sectionCommands.end();
1585 auto nonScriptI = std::find_if(i, e, [](SectionCommand *cmd) {
1586 if (auto *osd = dyn_cast<OutputDesc>(cmd))
1587 return osd->osec.sectionIndex == UINT32_MAX;
1588 return false;
1589 });
1590
1591 // Sort the orphan sections.
1592 std::stable_sort(nonScriptI, e, compareSections);
1593
1594 // As a horrible special case, skip the first . assignment if it is before any
1595 // section. We do this because it is common to set a load address by starting
1596 // the script with ". = 0xabcd" and the expectation is that every section is
1597 // after that.
1598 auto firstSectionOrDotAssignment =
1599 std::find_if(i, e, [](SectionCommand *cmd) { return !shouldSkip(cmd); });
1600 if (firstSectionOrDotAssignment != e &&
1601 isa<SymbolAssignment>(**firstSectionOrDotAssignment))
1602 ++firstSectionOrDotAssignment;
1603 i = firstSectionOrDotAssignment;
1604
1605 while (nonScriptI != e) {
1606 auto pos = findOrphanPos(i, nonScriptI);
1607 OutputSection *orphan = &cast<OutputDesc>(*nonScriptI)->osec;
1608
1609 // As an optimization, find all sections with the same sort rank
1610 // and insert them with one rotate.
1611 unsigned rank = orphan->sortRank;
1612 auto end = std::find_if(nonScriptI + 1, e, [=](SectionCommand *cmd) {
1613 return cast<OutputDesc>(cmd)->osec.sortRank != rank;
1614 });
1615 std::rotate(pos, nonScriptI, end);
1616 nonScriptI = end;
1617 }
1618 }
1619
compareByFilePosition(InputSection * a,InputSection * b)1620 static bool compareByFilePosition(InputSection *a, InputSection *b) {
1621 InputSection *la = a->flags & SHF_LINK_ORDER ? a->getLinkOrderDep() : nullptr;
1622 InputSection *lb = b->flags & SHF_LINK_ORDER ? b->getLinkOrderDep() : nullptr;
1623 // SHF_LINK_ORDER sections with non-zero sh_link are ordered before
1624 // non-SHF_LINK_ORDER sections and SHF_LINK_ORDER sections with zero sh_link.
1625 if (!la || !lb)
1626 return la && !lb;
1627 OutputSection *aOut = la->getParent();
1628 OutputSection *bOut = lb->getParent();
1629
1630 if (aOut != bOut)
1631 return aOut->addr < bOut->addr;
1632 return la->outSecOff < lb->outSecOff;
1633 }
1634
resolveShfLinkOrder()1635 template <class ELFT> void Writer<ELFT>::resolveShfLinkOrder() {
1636 llvm::TimeTraceScope timeScope("Resolve SHF_LINK_ORDER");
1637 for (OutputSection *sec : outputSections) {
1638 if (!(sec->flags & SHF_LINK_ORDER))
1639 continue;
1640
1641 // The ARM.exidx section use SHF_LINK_ORDER, but we have consolidated
1642 // this processing inside the ARMExidxsyntheticsection::finalizeContents().
1643 if (!config->relocatable && config->emachine == EM_ARM &&
1644 sec->type == SHT_ARM_EXIDX)
1645 continue;
1646
1647 // Link order may be distributed across several InputSectionDescriptions.
1648 // Sorting is performed separately.
1649 SmallVector<InputSection **, 0> scriptSections;
1650 SmallVector<InputSection *, 0> sections;
1651 for (SectionCommand *cmd : sec->commands) {
1652 auto *isd = dyn_cast<InputSectionDescription>(cmd);
1653 if (!isd)
1654 continue;
1655 bool hasLinkOrder = false;
1656 scriptSections.clear();
1657 sections.clear();
1658 for (InputSection *&isec : isd->sections) {
1659 if (isec->flags & SHF_LINK_ORDER) {
1660 InputSection *link = isec->getLinkOrderDep();
1661 if (link && !link->getParent())
1662 error(toString(isec) + ": sh_link points to discarded section " +
1663 toString(link));
1664 hasLinkOrder = true;
1665 }
1666 scriptSections.push_back(&isec);
1667 sections.push_back(isec);
1668 }
1669 if (hasLinkOrder && errorCount() == 0) {
1670 llvm::stable_sort(sections, compareByFilePosition);
1671 for (int i = 0, n = sections.size(); i != n; ++i)
1672 *scriptSections[i] = sections[i];
1673 }
1674 }
1675 }
1676 }
1677
finalizeSynthetic(SyntheticSection * sec)1678 static void finalizeSynthetic(SyntheticSection *sec) {
1679 if (sec && sec->isNeeded() && sec->getParent()) {
1680 llvm::TimeTraceScope timeScope("Finalize synthetic sections", sec->name);
1681 sec->finalizeContents();
1682 }
1683 }
1684
1685 // We need to generate and finalize the content that depends on the address of
1686 // InputSections. As the generation of the content may also alter InputSection
1687 // addresses we must converge to a fixed point. We do that here. See the comment
1688 // in Writer<ELFT>::finalizeSections().
finalizeAddressDependentContent()1689 template <class ELFT> void Writer<ELFT>::finalizeAddressDependentContent() {
1690 llvm::TimeTraceScope timeScope("Finalize address dependent content");
1691 ThunkCreator tc;
1692 AArch64Err843419Patcher a64p;
1693 ARMErr657417Patcher a32p;
1694 script->assignAddresses();
1695 // .ARM.exidx and SHF_LINK_ORDER do not require precise addresses, but they
1696 // do require the relative addresses of OutputSections because linker scripts
1697 // can assign Virtual Addresses to OutputSections that are not monotonically
1698 // increasing.
1699 for (Partition &part : partitions)
1700 finalizeSynthetic(part.armExidx.get());
1701 resolveShfLinkOrder();
1702
1703 // Converts call x@GDPLT to call __tls_get_addr
1704 if (config->emachine == EM_HEXAGON)
1705 hexagonTLSSymbolUpdate(outputSections);
1706
1707 uint32_t pass = 0, assignPasses = 0;
1708 for (;;) {
1709 bool changed = target->needsThunks ? tc.createThunks(pass, outputSections)
1710 : target->relaxOnce(pass);
1711 ++pass;
1712
1713 // With Thunk Size much smaller than branch range we expect to
1714 // converge quickly; if we get to 30 something has gone wrong.
1715 if (changed && pass >= 30) {
1716 error(target->needsThunks ? "thunk creation not converged"
1717 : "relaxation not converged");
1718 break;
1719 }
1720
1721 if (config->fixCortexA53Errata843419) {
1722 if (changed)
1723 script->assignAddresses();
1724 changed |= a64p.createFixes();
1725 }
1726 if (config->fixCortexA8) {
1727 if (changed)
1728 script->assignAddresses();
1729 changed |= a32p.createFixes();
1730 }
1731
1732 finalizeSynthetic(in.got.get());
1733 if (in.mipsGot)
1734 in.mipsGot->updateAllocSize();
1735
1736 for (Partition &part : partitions) {
1737 changed |= part.relaDyn->updateAllocSize();
1738 if (part.relrDyn)
1739 changed |= part.relrDyn->updateAllocSize();
1740 if (part.memtagGlobalDescriptors)
1741 changed |= part.memtagGlobalDescriptors->updateAllocSize();
1742 }
1743
1744 const Defined *changedSym = script->assignAddresses();
1745 if (!changed) {
1746 // Some symbols may be dependent on section addresses. When we break the
1747 // loop, the symbol values are finalized because a previous
1748 // assignAddresses() finalized section addresses.
1749 if (!changedSym)
1750 break;
1751 if (++assignPasses == 5) {
1752 errorOrWarn("assignment to symbol " + toString(*changedSym) +
1753 " does not converge");
1754 break;
1755 }
1756 }
1757 }
1758 if (!config->relocatable)
1759 target->finalizeRelax(pass);
1760
1761 if (config->relocatable)
1762 for (OutputSection *sec : outputSections)
1763 sec->addr = 0;
1764
1765 // If addrExpr is set, the address may not be a multiple of the alignment.
1766 // Warn because this is error-prone.
1767 for (SectionCommand *cmd : script->sectionCommands)
1768 if (auto *osd = dyn_cast<OutputDesc>(cmd)) {
1769 OutputSection *osec = &osd->osec;
1770 if (osec->addr % osec->addralign != 0)
1771 warn("address (0x" + Twine::utohexstr(osec->addr) + ") of section " +
1772 osec->name + " is not a multiple of alignment (" +
1773 Twine(osec->addralign) + ")");
1774 }
1775 }
1776
1777 // If Input Sections have been shrunk (basic block sections) then
1778 // update symbol values and sizes associated with these sections. With basic
1779 // block sections, input sections can shrink when the jump instructions at
1780 // the end of the section are relaxed.
fixSymbolsAfterShrinking()1781 static void fixSymbolsAfterShrinking() {
1782 for (InputFile *File : ctx.objectFiles) {
1783 parallelForEach(File->getSymbols(), [&](Symbol *Sym) {
1784 auto *def = dyn_cast<Defined>(Sym);
1785 if (!def)
1786 return;
1787
1788 const SectionBase *sec = def->section;
1789 if (!sec)
1790 return;
1791
1792 const InputSectionBase *inputSec = dyn_cast<InputSectionBase>(sec);
1793 if (!inputSec || !inputSec->bytesDropped)
1794 return;
1795
1796 const size_t OldSize = inputSec->content().size();
1797 const size_t NewSize = OldSize - inputSec->bytesDropped;
1798
1799 if (def->value > NewSize && def->value <= OldSize) {
1800 LLVM_DEBUG(llvm::dbgs()
1801 << "Moving symbol " << Sym->getName() << " from "
1802 << def->value << " to "
1803 << def->value - inputSec->bytesDropped << " bytes\n");
1804 def->value -= inputSec->bytesDropped;
1805 return;
1806 }
1807
1808 if (def->value + def->size > NewSize && def->value <= OldSize &&
1809 def->value + def->size <= OldSize) {
1810 LLVM_DEBUG(llvm::dbgs()
1811 << "Shrinking symbol " << Sym->getName() << " from "
1812 << def->size << " to " << def->size - inputSec->bytesDropped
1813 << " bytes\n");
1814 def->size -= inputSec->bytesDropped;
1815 }
1816 });
1817 }
1818 }
1819
1820 // If basic block sections exist, there are opportunities to delete fall thru
1821 // jumps and shrink jump instructions after basic block reordering. This
1822 // relaxation pass does that. It is only enabled when --optimize-bb-jumps
1823 // option is used.
optimizeBasicBlockJumps()1824 template <class ELFT> void Writer<ELFT>::optimizeBasicBlockJumps() {
1825 assert(config->optimizeBBJumps);
1826 SmallVector<InputSection *, 0> storage;
1827
1828 script->assignAddresses();
1829 // For every output section that has executable input sections, this
1830 // does the following:
1831 // 1. Deletes all direct jump instructions in input sections that
1832 // jump to the following section as it is not required.
1833 // 2. If there are two consecutive jump instructions, it checks
1834 // if they can be flipped and one can be deleted.
1835 for (OutputSection *osec : outputSections) {
1836 if (!(osec->flags & SHF_EXECINSTR))
1837 continue;
1838 ArrayRef<InputSection *> sections = getInputSections(*osec, storage);
1839 size_t numDeleted = 0;
1840 // Delete all fall through jump instructions. Also, check if two
1841 // consecutive jump instructions can be flipped so that a fall
1842 // through jmp instruction can be deleted.
1843 for (size_t i = 0, e = sections.size(); i != e; ++i) {
1844 InputSection *next = i + 1 < sections.size() ? sections[i + 1] : nullptr;
1845 InputSection &sec = *sections[i];
1846 numDeleted += target->deleteFallThruJmpInsn(sec, sec.file, next);
1847 }
1848 if (numDeleted > 0) {
1849 script->assignAddresses();
1850 LLVM_DEBUG(llvm::dbgs()
1851 << "Removing " << numDeleted << " fall through jumps\n");
1852 }
1853 }
1854
1855 fixSymbolsAfterShrinking();
1856
1857 for (OutputSection *osec : outputSections)
1858 for (InputSection *is : getInputSections(*osec, storage))
1859 is->trim();
1860 }
1861
1862 // In order to allow users to manipulate linker-synthesized sections,
1863 // we had to add synthetic sections to the input section list early,
1864 // even before we make decisions whether they are needed. This allows
1865 // users to write scripts like this: ".mygot : { .got }".
1866 //
1867 // Doing it has an unintended side effects. If it turns out that we
1868 // don't need a .got (for example) at all because there's no
1869 // relocation that needs a .got, we don't want to emit .got.
1870 //
1871 // To deal with the above problem, this function is called after
1872 // scanRelocations is called to remove synthetic sections that turn
1873 // out to be empty.
removeUnusedSyntheticSections()1874 static void removeUnusedSyntheticSections() {
1875 // All input synthetic sections that can be empty are placed after
1876 // all regular ones. Reverse iterate to find the first synthetic section
1877 // after a non-synthetic one which will be our starting point.
1878 auto start =
1879 llvm::find_if(llvm::reverse(ctx.inputSections), [](InputSectionBase *s) {
1880 return !isa<SyntheticSection>(s);
1881 }).base();
1882
1883 // Remove unused synthetic sections from ctx.inputSections;
1884 DenseSet<InputSectionBase *> unused;
1885 auto end =
1886 std::remove_if(start, ctx.inputSections.end(), [&](InputSectionBase *s) {
1887 auto *sec = cast<SyntheticSection>(s);
1888 if (sec->getParent() && sec->isNeeded())
1889 return false;
1890 unused.insert(sec);
1891 return true;
1892 });
1893 ctx.inputSections.erase(end, ctx.inputSections.end());
1894
1895 // Remove unused synthetic sections from the corresponding input section
1896 // description and orphanSections.
1897 for (auto *sec : unused)
1898 if (OutputSection *osec = cast<SyntheticSection>(sec)->getParent())
1899 for (SectionCommand *cmd : osec->commands)
1900 if (auto *isd = dyn_cast<InputSectionDescription>(cmd))
1901 llvm::erase_if(isd->sections, [&](InputSection *isec) {
1902 return unused.count(isec);
1903 });
1904 llvm::erase_if(script->orphanSections, [&](const InputSectionBase *sec) {
1905 return unused.count(sec);
1906 });
1907 }
1908
1909 // Create output section objects and add them to OutputSections.
finalizeSections()1910 template <class ELFT> void Writer<ELFT>::finalizeSections() {
1911 if (!config->relocatable) {
1912 Out::preinitArray = findSection(".preinit_array");
1913 Out::initArray = findSection(".init_array");
1914 Out::finiArray = findSection(".fini_array");
1915
1916 // The linker needs to define SECNAME_start, SECNAME_end and SECNAME_stop
1917 // symbols for sections, so that the runtime can get the start and end
1918 // addresses of each section by section name. Add such symbols.
1919 addStartEndSymbols();
1920 for (SectionCommand *cmd : script->sectionCommands)
1921 if (auto *osd = dyn_cast<OutputDesc>(cmd))
1922 addStartStopSymbols(osd->osec);
1923
1924 // Add _DYNAMIC symbol. Unlike GNU gold, our _DYNAMIC symbol has no type.
1925 // It should be okay as no one seems to care about the type.
1926 // Even the author of gold doesn't remember why gold behaves that way.
1927 // https://sourceware.org/ml/binutils/2002-03/msg00360.html
1928 if (mainPart->dynamic->parent) {
1929 Symbol *s = symtab.addSymbol(Defined{
1930 ctx.internalFile, "_DYNAMIC", STB_WEAK, STV_HIDDEN, STT_NOTYPE,
1931 /*value=*/0, /*size=*/0, mainPart->dynamic.get()});
1932 s->isUsedInRegularObj = true;
1933 }
1934
1935 // Define __rel[a]_iplt_{start,end} symbols if needed.
1936 addRelIpltSymbols();
1937
1938 // RISC-V's gp can address +/- 2 KiB, set it to .sdata + 0x800. This symbol
1939 // should only be defined in an executable. If .sdata does not exist, its
1940 // value/section does not matter but it has to be relative, so set its
1941 // st_shndx arbitrarily to 1 (Out::elfHeader).
1942 if (config->emachine == EM_RISCV) {
1943 ElfSym::riscvGlobalPointer = nullptr;
1944 if (!config->shared) {
1945 OutputSection *sec = findSection(".sdata");
1946 addOptionalRegular(
1947 "__global_pointer$", sec ? sec : Out::elfHeader, 0x800, STV_DEFAULT);
1948 // Set riscvGlobalPointer to be used by the optional global pointer
1949 // relaxation.
1950 if (config->relaxGP) {
1951 Symbol *s = symtab.find("__global_pointer$");
1952 if (s && s->isDefined())
1953 ElfSym::riscvGlobalPointer = cast<Defined>(s);
1954 }
1955 }
1956 }
1957
1958 if (config->emachine == EM_386 || config->emachine == EM_X86_64) {
1959 // On targets that support TLSDESC, _TLS_MODULE_BASE_ is defined in such a
1960 // way that:
1961 //
1962 // 1) Without relaxation: it produces a dynamic TLSDESC relocation that
1963 // computes 0.
1964 // 2) With LD->LE relaxation: _TLS_MODULE_BASE_@tpoff = 0 (lowest address
1965 // in the TLS block).
1966 //
1967 // 2) is special cased in @tpoff computation. To satisfy 1), we define it
1968 // as an absolute symbol of zero. This is different from GNU linkers which
1969 // define _TLS_MODULE_BASE_ relative to the first TLS section.
1970 Symbol *s = symtab.find("_TLS_MODULE_BASE_");
1971 if (s && s->isUndefined()) {
1972 s->resolve(Defined{ctx.internalFile, StringRef(), STB_GLOBAL,
1973 STV_HIDDEN, STT_TLS, /*value=*/0, 0,
1974 /*section=*/nullptr});
1975 ElfSym::tlsModuleBase = cast<Defined>(s);
1976 }
1977 }
1978
1979 // This responsible for splitting up .eh_frame section into
1980 // pieces. The relocation scan uses those pieces, so this has to be
1981 // earlier.
1982 {
1983 llvm::TimeTraceScope timeScope("Finalize .eh_frame");
1984 for (Partition &part : partitions)
1985 finalizeSynthetic(part.ehFrame.get());
1986 }
1987 }
1988
1989 demoteSymbolsAndComputeIsPreemptible();
1990
1991 if (config->copyRelocs && config->discard != DiscardPolicy::None)
1992 markUsedLocalSymbols<ELFT>();
1993 demoteAndCopyLocalSymbols();
1994
1995 if (config->copyRelocs)
1996 addSectionSymbols();
1997
1998 // Change values of linker-script-defined symbols from placeholders (assigned
1999 // by declareSymbols) to actual definitions.
2000 script->processSymbolAssignments();
2001
2002 if (!config->relocatable) {
2003 llvm::TimeTraceScope timeScope("Scan relocations");
2004 // Scan relocations. This must be done after every symbol is declared so
2005 // that we can correctly decide if a dynamic relocation is needed. This is
2006 // called after processSymbolAssignments() because it needs to know whether
2007 // a linker-script-defined symbol is absolute.
2008 ppc64noTocRelax.clear();
2009 scanRelocations<ELFT>();
2010 reportUndefinedSymbols();
2011 postScanRelocations();
2012
2013 if (in.plt && in.plt->isNeeded())
2014 in.plt->addSymbols();
2015 if (in.iplt && in.iplt->isNeeded())
2016 in.iplt->addSymbols();
2017
2018 if (config->unresolvedSymbolsInShlib != UnresolvedPolicy::Ignore) {
2019 auto diagnose =
2020 config->unresolvedSymbolsInShlib == UnresolvedPolicy::ReportError
2021 ? errorOrWarn
2022 : warn;
2023 // Error on undefined symbols in a shared object, if all of its DT_NEEDED
2024 // entries are seen. These cases would otherwise lead to runtime errors
2025 // reported by the dynamic linker.
2026 //
2027 // ld.bfd traces all DT_NEEDED to emulate the logic of the dynamic linker
2028 // to catch more cases. That is too much for us. Our approach resembles
2029 // the one used in ld.gold, achieves a good balance to be useful but not
2030 // too smart.
2031 //
2032 // If a DSO reference is resolved by a SharedSymbol, but the SharedSymbol
2033 // is overridden by a hidden visibility Defined (which is later discarded
2034 // due to GC), don't report the diagnostic. However, this may indicate an
2035 // unintended SharedSymbol.
2036 for (SharedFile *file : ctx.sharedFiles) {
2037 bool allNeededIsKnown =
2038 llvm::all_of(file->dtNeeded, [&](StringRef needed) {
2039 return symtab.soNames.count(CachedHashStringRef(needed));
2040 });
2041 if (!allNeededIsKnown)
2042 continue;
2043 for (Symbol *sym : file->requiredSymbols) {
2044 if (sym->dsoDefined)
2045 continue;
2046 if (sym->isUndefined() && !sym->isWeak()) {
2047 diagnose("undefined reference due to --no-allow-shlib-undefined: " +
2048 toString(*sym) + "\n>>> referenced by " + toString(file));
2049 } else if (sym->isDefined() && sym->computeBinding() == STB_LOCAL) {
2050 diagnose("non-exported symbol '" + toString(*sym) + "' in '" +
2051 toString(sym->file) + "' is referenced by DSO '" +
2052 toString(file) + "'");
2053 }
2054 }
2055 }
2056 }
2057 }
2058
2059 {
2060 llvm::TimeTraceScope timeScope("Add symbols to symtabs");
2061 // Now that we have defined all possible global symbols including linker-
2062 // synthesized ones. Visit all symbols to give the finishing touches.
2063 for (Symbol *sym : symtab.getSymbols()) {
2064 if (!sym->isUsedInRegularObj || !includeInSymtab(*sym))
2065 continue;
2066 if (!config->relocatable)
2067 sym->binding = sym->computeBinding();
2068 if (in.symTab)
2069 in.symTab->addSymbol(sym);
2070
2071 if (sym->includeInDynsym()) {
2072 partitions[sym->partition - 1].dynSymTab->addSymbol(sym);
2073 if (auto *file = dyn_cast_or_null<SharedFile>(sym->file))
2074 if (file->isNeeded && !sym->isUndefined())
2075 addVerneed(sym);
2076 }
2077 }
2078
2079 // We also need to scan the dynamic relocation tables of the other
2080 // partitions and add any referenced symbols to the partition's dynsym.
2081 for (Partition &part : MutableArrayRef<Partition>(partitions).slice(1)) {
2082 DenseSet<Symbol *> syms;
2083 for (const SymbolTableEntry &e : part.dynSymTab->getSymbols())
2084 syms.insert(e.sym);
2085 for (DynamicReloc &reloc : part.relaDyn->relocs)
2086 if (reloc.sym && reloc.needsDynSymIndex() &&
2087 syms.insert(reloc.sym).second)
2088 part.dynSymTab->addSymbol(reloc.sym);
2089 }
2090 }
2091
2092 if (in.mipsGot)
2093 in.mipsGot->build();
2094
2095 removeUnusedSyntheticSections();
2096 script->diagnoseOrphanHandling();
2097 script->diagnoseMissingSGSectionAddress();
2098
2099 sortSections();
2100
2101 // Create a list of OutputSections, assign sectionIndex, and populate
2102 // in.shStrTab.
2103 for (SectionCommand *cmd : script->sectionCommands)
2104 if (auto *osd = dyn_cast<OutputDesc>(cmd)) {
2105 OutputSection *osec = &osd->osec;
2106 outputSections.push_back(osec);
2107 osec->sectionIndex = outputSections.size();
2108 osec->shName = in.shStrTab->addString(osec->name);
2109 }
2110
2111 // Prefer command line supplied address over other constraints.
2112 for (OutputSection *sec : outputSections) {
2113 auto i = config->sectionStartMap.find(sec->name);
2114 if (i != config->sectionStartMap.end())
2115 sec->addrExpr = [=] { return i->second; };
2116 }
2117
2118 // With the outputSections available check for GDPLT relocations
2119 // and add __tls_get_addr symbol if needed.
2120 if (config->emachine == EM_HEXAGON && hexagonNeedsTLSSymbol(outputSections)) {
2121 Symbol *sym =
2122 symtab.addSymbol(Undefined{ctx.internalFile, "__tls_get_addr",
2123 STB_GLOBAL, STV_DEFAULT, STT_NOTYPE});
2124 sym->isPreemptible = true;
2125 partitions[0].dynSymTab->addSymbol(sym);
2126 }
2127
2128 // This is a bit of a hack. A value of 0 means undef, so we set it
2129 // to 1 to make __ehdr_start defined. The section number is not
2130 // particularly relevant.
2131 Out::elfHeader->sectionIndex = 1;
2132 Out::elfHeader->size = sizeof(typename ELFT::Ehdr);
2133
2134 // Binary and relocatable output does not have PHDRS.
2135 // The headers have to be created before finalize as that can influence the
2136 // image base and the dynamic section on mips includes the image base.
2137 if (!config->relocatable && !config->oFormatBinary) {
2138 for (Partition &part : partitions) {
2139 part.phdrs = script->hasPhdrsCommands() ? script->createPhdrs()
2140 : createPhdrs(part);
2141 if (config->emachine == EM_ARM) {
2142 // PT_ARM_EXIDX is the ARM EHABI equivalent of PT_GNU_EH_FRAME
2143 addPhdrForSection(part, SHT_ARM_EXIDX, PT_ARM_EXIDX, PF_R);
2144 }
2145 if (config->emachine == EM_MIPS) {
2146 // Add separate segments for MIPS-specific sections.
2147 addPhdrForSection(part, SHT_MIPS_REGINFO, PT_MIPS_REGINFO, PF_R);
2148 addPhdrForSection(part, SHT_MIPS_OPTIONS, PT_MIPS_OPTIONS, PF_R);
2149 addPhdrForSection(part, SHT_MIPS_ABIFLAGS, PT_MIPS_ABIFLAGS, PF_R);
2150 }
2151 if (config->emachine == EM_RISCV)
2152 addPhdrForSection(part, SHT_RISCV_ATTRIBUTES, PT_RISCV_ATTRIBUTES,
2153 PF_R);
2154 }
2155 Out::programHeaders->size = sizeof(Elf_Phdr) * mainPart->phdrs.size();
2156
2157 // Find the TLS segment. This happens before the section layout loop so that
2158 // Android relocation packing can look up TLS symbol addresses. We only need
2159 // to care about the main partition here because all TLS symbols were moved
2160 // to the main partition (see MarkLive.cpp).
2161 for (PhdrEntry *p : mainPart->phdrs)
2162 if (p->p_type == PT_TLS)
2163 Out::tlsPhdr = p;
2164 }
2165
2166 // Some symbols are defined in term of program headers. Now that we
2167 // have the headers, we can find out which sections they point to.
2168 setReservedSymbolSections();
2169
2170 {
2171 llvm::TimeTraceScope timeScope("Finalize synthetic sections");
2172
2173 finalizeSynthetic(in.bss.get());
2174 finalizeSynthetic(in.bssRelRo.get());
2175 finalizeSynthetic(in.symTabShndx.get());
2176 finalizeSynthetic(in.shStrTab.get());
2177 finalizeSynthetic(in.strTab.get());
2178 finalizeSynthetic(in.got.get());
2179 finalizeSynthetic(in.mipsGot.get());
2180 finalizeSynthetic(in.igotPlt.get());
2181 finalizeSynthetic(in.gotPlt.get());
2182 finalizeSynthetic(in.relaIplt.get());
2183 finalizeSynthetic(in.relaPlt.get());
2184 finalizeSynthetic(in.plt.get());
2185 finalizeSynthetic(in.iplt.get());
2186 finalizeSynthetic(in.ppc32Got2.get());
2187 finalizeSynthetic(in.partIndex.get());
2188
2189 // Dynamic section must be the last one in this list and dynamic
2190 // symbol table section (dynSymTab) must be the first one.
2191 for (Partition &part : partitions) {
2192 if (part.relaDyn) {
2193 part.relaDyn->mergeRels();
2194 // Compute DT_RELACOUNT to be used by part.dynamic.
2195 part.relaDyn->partitionRels();
2196 finalizeSynthetic(part.relaDyn.get());
2197 }
2198 if (part.relrDyn) {
2199 part.relrDyn->mergeRels();
2200 finalizeSynthetic(part.relrDyn.get());
2201 }
2202
2203 finalizeSynthetic(part.dynSymTab.get());
2204 finalizeSynthetic(part.gnuHashTab.get());
2205 finalizeSynthetic(part.hashTab.get());
2206 finalizeSynthetic(part.verDef.get());
2207 finalizeSynthetic(part.ehFrameHdr.get());
2208 finalizeSynthetic(part.verSym.get());
2209 finalizeSynthetic(part.verNeed.get());
2210 finalizeSynthetic(part.dynamic.get());
2211 }
2212 }
2213
2214 if (!script->hasSectionsCommand && !config->relocatable)
2215 fixSectionAlignments();
2216
2217 // This is used to:
2218 // 1) Create "thunks":
2219 // Jump instructions in many ISAs have small displacements, and therefore
2220 // they cannot jump to arbitrary addresses in memory. For example, RISC-V
2221 // JAL instruction can target only +-1 MiB from PC. It is a linker's
2222 // responsibility to create and insert small pieces of code between
2223 // sections to extend the ranges if jump targets are out of range. Such
2224 // code pieces are called "thunks".
2225 //
2226 // We add thunks at this stage. We couldn't do this before this point
2227 // because this is the earliest point where we know sizes of sections and
2228 // their layouts (that are needed to determine if jump targets are in
2229 // range).
2230 //
2231 // 2) Update the sections. We need to generate content that depends on the
2232 // address of InputSections. For example, MIPS GOT section content or
2233 // android packed relocations sections content.
2234 //
2235 // 3) Assign the final values for the linker script symbols. Linker scripts
2236 // sometimes using forward symbol declarations. We want to set the correct
2237 // values. They also might change after adding the thunks.
2238 finalizeAddressDependentContent();
2239
2240 // All information needed for OutputSection part of Map file is available.
2241 if (errorCount())
2242 return;
2243
2244 {
2245 llvm::TimeTraceScope timeScope("Finalize synthetic sections");
2246 // finalizeAddressDependentContent may have added local symbols to the
2247 // static symbol table.
2248 finalizeSynthetic(in.symTab.get());
2249 finalizeSynthetic(in.ppc64LongBranchTarget.get());
2250 finalizeSynthetic(in.armCmseSGSection.get());
2251 }
2252
2253 // Relaxation to delete inter-basic block jumps created by basic block
2254 // sections. Run after in.symTab is finalized as optimizeBasicBlockJumps
2255 // can relax jump instructions based on symbol offset.
2256 if (config->optimizeBBJumps)
2257 optimizeBasicBlockJumps();
2258
2259 // Fill other section headers. The dynamic table is finalized
2260 // at the end because some tags like RELSZ depend on result
2261 // of finalizing other sections.
2262 for (OutputSection *sec : outputSections)
2263 sec->finalize();
2264
2265 script->checkFinalScriptConditions();
2266
2267 if (config->emachine == EM_ARM && !config->isLE && config->armBe8) {
2268 addArmInputSectionMappingSymbols();
2269 sortArmMappingSymbols();
2270 }
2271 }
2272
2273 // Ensure data sections are not mixed with executable sections when
2274 // --execute-only is used. --execute-only make pages executable but not
2275 // readable.
checkExecuteOnly()2276 template <class ELFT> void Writer<ELFT>::checkExecuteOnly() {
2277 if (!config->executeOnly)
2278 return;
2279
2280 SmallVector<InputSection *, 0> storage;
2281 for (OutputSection *osec : outputSections)
2282 if (osec->flags & SHF_EXECINSTR)
2283 for (InputSection *isec : getInputSections(*osec, storage))
2284 if (!(isec->flags & SHF_EXECINSTR))
2285 error("cannot place " + toString(isec) + " into " +
2286 toString(osec->name) +
2287 ": --execute-only does not support intermingling data and code");
2288 }
2289
2290 // The linker is expected to define SECNAME_start and SECNAME_end
2291 // symbols for a few sections. This function defines them.
addStartEndSymbols()2292 template <class ELFT> void Writer<ELFT>::addStartEndSymbols() {
2293 // If a section does not exist, there's ambiguity as to how we
2294 // define _start and _end symbols for an init/fini section. Since
2295 // the loader assume that the symbols are always defined, we need to
2296 // always define them. But what value? The loader iterates over all
2297 // pointers between _start and _end to run global ctors/dtors, so if
2298 // the section is empty, their symbol values don't actually matter
2299 // as long as _start and _end point to the same location.
2300 //
2301 // That said, we don't want to set the symbols to 0 (which is
2302 // probably the simplest value) because that could cause some
2303 // program to fail to link due to relocation overflow, if their
2304 // program text is above 2 GiB. We use the address of the .text
2305 // section instead to prevent that failure.
2306 //
2307 // In rare situations, the .text section may not exist. If that's the
2308 // case, use the image base address as a last resort.
2309 OutputSection *Default = findSection(".text");
2310 if (!Default)
2311 Default = Out::elfHeader;
2312
2313 auto define = [=](StringRef start, StringRef end, OutputSection *os) {
2314 if (os && !script->isDiscarded(os)) {
2315 addOptionalRegular(start, os, 0);
2316 addOptionalRegular(end, os, -1);
2317 } else {
2318 addOptionalRegular(start, Default, 0);
2319 addOptionalRegular(end, Default, 0);
2320 }
2321 };
2322
2323 define("__preinit_array_start", "__preinit_array_end", Out::preinitArray);
2324 define("__init_array_start", "__init_array_end", Out::initArray);
2325 define("__fini_array_start", "__fini_array_end", Out::finiArray);
2326
2327 if (OutputSection *sec = findSection(".ARM.exidx"))
2328 define("__exidx_start", "__exidx_end", sec);
2329 }
2330
2331 // If a section name is valid as a C identifier (which is rare because of
2332 // the leading '.'), linkers are expected to define __start_<secname> and
2333 // __stop_<secname> symbols. They are at beginning and end of the section,
2334 // respectively. This is not requested by the ELF standard, but GNU ld and
2335 // gold provide the feature, and used by many programs.
2336 template <class ELFT>
addStartStopSymbols(OutputSection & osec)2337 void Writer<ELFT>::addStartStopSymbols(OutputSection &osec) {
2338 StringRef s = osec.name;
2339 if (!isValidCIdentifier(s))
2340 return;
2341 addOptionalRegular(saver().save("__start_" + s), &osec, 0,
2342 config->zStartStopVisibility);
2343 addOptionalRegular(saver().save("__stop_" + s), &osec, -1,
2344 config->zStartStopVisibility);
2345 }
2346
needsPtLoad(OutputSection * sec)2347 static bool needsPtLoad(OutputSection *sec) {
2348 if (!(sec->flags & SHF_ALLOC))
2349 return false;
2350
2351 // Don't allocate VA space for TLS NOBITS sections. The PT_TLS PHDR is
2352 // responsible for allocating space for them, not the PT_LOAD that
2353 // contains the TLS initialization image.
2354 if ((sec->flags & SHF_TLS) && sec->type == SHT_NOBITS)
2355 return false;
2356 return true;
2357 }
2358
2359 // Linker scripts are responsible for aligning addresses. Unfortunately, most
2360 // linker scripts are designed for creating two PT_LOADs only, one RX and one
2361 // RW. This means that there is no alignment in the RO to RX transition and we
2362 // cannot create a PT_LOAD there.
computeFlags(uint64_t flags)2363 static uint64_t computeFlags(uint64_t flags) {
2364 if (config->omagic)
2365 return PF_R | PF_W | PF_X;
2366 if (config->executeOnly && (flags & PF_X))
2367 return flags & ~PF_R;
2368 if (config->singleRoRx && !(flags & PF_W))
2369 return flags | PF_X;
2370 return flags;
2371 }
2372
2373 // Decide which program headers to create and which sections to include in each
2374 // one.
2375 template <class ELFT>
createPhdrs(Partition & part)2376 SmallVector<PhdrEntry *, 0> Writer<ELFT>::createPhdrs(Partition &part) {
2377 SmallVector<PhdrEntry *, 0> ret;
2378 auto addHdr = [&](unsigned type, unsigned flags) -> PhdrEntry * {
2379 ret.push_back(make<PhdrEntry>(type, flags));
2380 return ret.back();
2381 };
2382
2383 unsigned partNo = part.getNumber();
2384 bool isMain = partNo == 1;
2385
2386 // Add the first PT_LOAD segment for regular output sections.
2387 uint64_t flags = computeFlags(PF_R);
2388 PhdrEntry *load = nullptr;
2389
2390 // nmagic or omagic output does not have PT_PHDR, PT_INTERP, or the readonly
2391 // PT_LOAD.
2392 if (!config->nmagic && !config->omagic) {
2393 // The first phdr entry is PT_PHDR which describes the program header
2394 // itself.
2395 if (isMain)
2396 addHdr(PT_PHDR, PF_R)->add(Out::programHeaders);
2397 else
2398 addHdr(PT_PHDR, PF_R)->add(part.programHeaders->getParent());
2399
2400 // PT_INTERP must be the second entry if exists.
2401 if (OutputSection *cmd = findSection(".interp", partNo))
2402 addHdr(PT_INTERP, cmd->getPhdrFlags())->add(cmd);
2403
2404 // Add the headers. We will remove them if they don't fit.
2405 // In the other partitions the headers are ordinary sections, so they don't
2406 // need to be added here.
2407 if (isMain) {
2408 load = addHdr(PT_LOAD, flags);
2409 load->add(Out::elfHeader);
2410 load->add(Out::programHeaders);
2411 }
2412 }
2413
2414 // PT_GNU_RELRO includes all sections that should be marked as
2415 // read-only by dynamic linker after processing relocations.
2416 // Current dynamic loaders only support one PT_GNU_RELRO PHDR, give
2417 // an error message if more than one PT_GNU_RELRO PHDR is required.
2418 PhdrEntry *relRo = make<PhdrEntry>(PT_GNU_RELRO, PF_R);
2419 bool inRelroPhdr = false;
2420 OutputSection *relroEnd = nullptr;
2421 for (OutputSection *sec : outputSections) {
2422 if (sec->partition != partNo || !needsPtLoad(sec))
2423 continue;
2424 if (isRelroSection(sec)) {
2425 inRelroPhdr = true;
2426 if (!relroEnd)
2427 relRo->add(sec);
2428 else
2429 error("section: " + sec->name + " is not contiguous with other relro" +
2430 " sections");
2431 } else if (inRelroPhdr) {
2432 inRelroPhdr = false;
2433 relroEnd = sec;
2434 }
2435 }
2436 relRo->p_align = 1;
2437
2438 for (OutputSection *sec : outputSections) {
2439 if (!needsPtLoad(sec))
2440 continue;
2441
2442 // Normally, sections in partitions other than the current partition are
2443 // ignored. But partition number 255 is a special case: it contains the
2444 // partition end marker (.part.end). It needs to be added to the main
2445 // partition so that a segment is created for it in the main partition,
2446 // which will cause the dynamic loader to reserve space for the other
2447 // partitions.
2448 if (sec->partition != partNo) {
2449 if (isMain && sec->partition == 255)
2450 addHdr(PT_LOAD, computeFlags(sec->getPhdrFlags()))->add(sec);
2451 continue;
2452 }
2453
2454 // Segments are contiguous memory regions that has the same attributes
2455 // (e.g. executable or writable). There is one phdr for each segment.
2456 // Therefore, we need to create a new phdr when the next section has
2457 // different flags or is loaded at a discontiguous address or memory region
2458 // using AT or AT> linker script command, respectively.
2459 //
2460 // As an exception, we don't create a separate load segment for the ELF
2461 // headers, even if the first "real" output has an AT or AT> attribute.
2462 //
2463 // In addition, NOBITS sections should only be placed at the end of a LOAD
2464 // segment (since it's represented as p_filesz < p_memsz). If we have a
2465 // not-NOBITS section after a NOBITS, we create a new LOAD for the latter
2466 // even if flags match, so as not to require actually writing the
2467 // supposed-to-be-NOBITS section to the output file. (However, we cannot do
2468 // so when hasSectionsCommand, since we cannot introduce the extra alignment
2469 // needed to create a new LOAD)
2470 uint64_t newFlags = computeFlags(sec->getPhdrFlags());
2471 bool sameLMARegion =
2472 load && !sec->lmaExpr && sec->lmaRegion == load->firstSec->lmaRegion;
2473 if (!(load && newFlags == flags && sec != relroEnd &&
2474 sec->memRegion == load->firstSec->memRegion &&
2475 (sameLMARegion || load->lastSec == Out::programHeaders) &&
2476 (script->hasSectionsCommand || sec->type == SHT_NOBITS ||
2477 load->lastSec->type != SHT_NOBITS))) {
2478 load = addHdr(PT_LOAD, newFlags);
2479 flags = newFlags;
2480 }
2481
2482 load->add(sec);
2483 }
2484
2485 // Add a TLS segment if any.
2486 PhdrEntry *tlsHdr = make<PhdrEntry>(PT_TLS, PF_R);
2487 for (OutputSection *sec : outputSections)
2488 if (sec->partition == partNo && sec->flags & SHF_TLS)
2489 tlsHdr->add(sec);
2490 if (tlsHdr->firstSec)
2491 ret.push_back(tlsHdr);
2492
2493 // Add an entry for .dynamic.
2494 if (OutputSection *sec = part.dynamic->getParent())
2495 addHdr(PT_DYNAMIC, sec->getPhdrFlags())->add(sec);
2496
2497 if (relRo->firstSec)
2498 ret.push_back(relRo);
2499
2500 // PT_GNU_EH_FRAME is a special section pointing on .eh_frame_hdr.
2501 if (part.ehFrame->isNeeded() && part.ehFrameHdr &&
2502 part.ehFrame->getParent() && part.ehFrameHdr->getParent())
2503 addHdr(PT_GNU_EH_FRAME, part.ehFrameHdr->getParent()->getPhdrFlags())
2504 ->add(part.ehFrameHdr->getParent());
2505
2506 // PT_OPENBSD_RANDOMIZE is an OpenBSD-specific feature. That makes
2507 // the dynamic linker fill the segment with random data.
2508 if (OutputSection *cmd = findSection(".openbsd.randomdata", partNo))
2509 addHdr(PT_OPENBSD_RANDOMIZE, cmd->getPhdrFlags())->add(cmd);
2510
2511 if (config->zGnustack != GnuStackKind::None) {
2512 // PT_GNU_STACK is a special section to tell the loader to make the
2513 // pages for the stack non-executable. If you really want an executable
2514 // stack, you can pass -z execstack, but that's not recommended for
2515 // security reasons.
2516 unsigned perm = PF_R | PF_W;
2517 if (config->zGnustack == GnuStackKind::Exec)
2518 perm |= PF_X;
2519 addHdr(PT_GNU_STACK, perm)->p_memsz = config->zStackSize;
2520 }
2521
2522 // PT_OPENBSD_WXNEEDED is a OpenBSD-specific header to mark the executable
2523 // is expected to perform W^X violations, such as calling mprotect(2) or
2524 // mmap(2) with PROT_WRITE | PROT_EXEC, which is prohibited by default on
2525 // OpenBSD.
2526 if (config->zWxneeded)
2527 addHdr(PT_OPENBSD_WXNEEDED, PF_X);
2528
2529 if (OutputSection *cmd = findSection(".note.gnu.property", partNo))
2530 addHdr(PT_GNU_PROPERTY, PF_R)->add(cmd);
2531
2532 // Create one PT_NOTE per a group of contiguous SHT_NOTE sections with the
2533 // same alignment.
2534 PhdrEntry *note = nullptr;
2535 for (OutputSection *sec : outputSections) {
2536 if (sec->partition != partNo)
2537 continue;
2538 if (sec->type == SHT_NOTE && (sec->flags & SHF_ALLOC)) {
2539 if (!note || sec->lmaExpr || note->lastSec->addralign != sec->addralign)
2540 note = addHdr(PT_NOTE, PF_R);
2541 note->add(sec);
2542 } else {
2543 note = nullptr;
2544 }
2545 }
2546 return ret;
2547 }
2548
2549 template <class ELFT>
addPhdrForSection(Partition & part,unsigned shType,unsigned pType,unsigned pFlags)2550 void Writer<ELFT>::addPhdrForSection(Partition &part, unsigned shType,
2551 unsigned pType, unsigned pFlags) {
2552 unsigned partNo = part.getNumber();
2553 auto i = llvm::find_if(outputSections, [=](OutputSection *cmd) {
2554 return cmd->partition == partNo && cmd->type == shType;
2555 });
2556 if (i == outputSections.end())
2557 return;
2558
2559 PhdrEntry *entry = make<PhdrEntry>(pType, pFlags);
2560 entry->add(*i);
2561 part.phdrs.push_back(entry);
2562 }
2563
2564 // Place the first section of each PT_LOAD to a different page (of maxPageSize).
2565 // This is achieved by assigning an alignment expression to addrExpr of each
2566 // such section.
fixSectionAlignments()2567 template <class ELFT> void Writer<ELFT>::fixSectionAlignments() {
2568 const PhdrEntry *prev;
2569 auto pageAlign = [&](const PhdrEntry *p) {
2570 OutputSection *cmd = p->firstSec;
2571 if (!cmd)
2572 return;
2573 cmd->alignExpr = [align = cmd->addralign]() { return align; };
2574 if (!cmd->addrExpr) {
2575 // Prefer advancing to align(dot, maxPageSize) + dot%maxPageSize to avoid
2576 // padding in the file contents.
2577 //
2578 // When -z separate-code is used we must not have any overlap in pages
2579 // between an executable segment and a non-executable segment. We align to
2580 // the next maximum page size boundary on transitions between executable
2581 // and non-executable segments.
2582 //
2583 // SHT_LLVM_PART_EHDR marks the start of a partition. The partition
2584 // sections will be extracted to a separate file. Align to the next
2585 // maximum page size boundary so that we can find the ELF header at the
2586 // start. We cannot benefit from overlapping p_offset ranges with the
2587 // previous segment anyway.
2588 if (config->zSeparate == SeparateSegmentKind::Loadable ||
2589 (config->zSeparate == SeparateSegmentKind::Code && prev &&
2590 (prev->p_flags & PF_X) != (p->p_flags & PF_X)) ||
2591 cmd->type == SHT_LLVM_PART_EHDR)
2592 cmd->addrExpr = [] {
2593 return alignToPowerOf2(script->getDot(), config->maxPageSize);
2594 };
2595 // PT_TLS is at the start of the first RW PT_LOAD. If `p` includes PT_TLS,
2596 // it must be the RW. Align to p_align(PT_TLS) to make sure
2597 // p_vaddr(PT_LOAD)%p_align(PT_LOAD) = 0. Otherwise, if
2598 // sh_addralign(.tdata) < sh_addralign(.tbss), we will set p_align(PT_TLS)
2599 // to sh_addralign(.tbss), while p_vaddr(PT_TLS)=p_vaddr(PT_LOAD) may not
2600 // be congruent to 0 modulo p_align(PT_TLS).
2601 //
2602 // Technically this is not required, but as of 2019, some dynamic loaders
2603 // don't handle p_vaddr%p_align != 0 correctly, e.g. glibc (i386 and
2604 // x86-64) doesn't make runtime address congruent to p_vaddr modulo
2605 // p_align for dynamic TLS blocks (PR/24606), FreeBSD rtld has the same
2606 // bug, musl (TLS Variant 1 architectures) before 1.1.23 handled TLS
2607 // blocks correctly. We need to keep the workaround for a while.
2608 else if (Out::tlsPhdr && Out::tlsPhdr->firstSec == p->firstSec)
2609 cmd->addrExpr = [] {
2610 return alignToPowerOf2(script->getDot(), config->maxPageSize) +
2611 alignToPowerOf2(script->getDot() % config->maxPageSize,
2612 Out::tlsPhdr->p_align);
2613 };
2614 else
2615 cmd->addrExpr = [] {
2616 return alignToPowerOf2(script->getDot(), config->maxPageSize) +
2617 script->getDot() % config->maxPageSize;
2618 };
2619 }
2620 };
2621
2622 for (Partition &part : partitions) {
2623 prev = nullptr;
2624 for (const PhdrEntry *p : part.phdrs)
2625 if (p->p_type == PT_LOAD && p->firstSec) {
2626 pageAlign(p);
2627 prev = p;
2628 }
2629 }
2630 }
2631
2632 // Compute an in-file position for a given section. The file offset must be the
2633 // same with its virtual address modulo the page size, so that the loader can
2634 // load executables without any address adjustment.
computeFileOffset(OutputSection * os,uint64_t off)2635 static uint64_t computeFileOffset(OutputSection *os, uint64_t off) {
2636 // The first section in a PT_LOAD has to have congruent offset and address
2637 // modulo the maximum page size.
2638 if (os->ptLoad && os->ptLoad->firstSec == os)
2639 return alignTo(off, os->ptLoad->p_align, os->addr);
2640
2641 // File offsets are not significant for .bss sections other than the first one
2642 // in a PT_LOAD/PT_TLS. By convention, we keep section offsets monotonically
2643 // increasing rather than setting to zero.
2644 if (os->type == SHT_NOBITS &&
2645 (!Out::tlsPhdr || Out::tlsPhdr->firstSec != os))
2646 return off;
2647
2648 // If the section is not in a PT_LOAD, we just have to align it.
2649 if (!os->ptLoad)
2650 return alignToPowerOf2(off, os->addralign);
2651
2652 // If two sections share the same PT_LOAD the file offset is calculated
2653 // using this formula: Off2 = Off1 + (VA2 - VA1).
2654 OutputSection *first = os->ptLoad->firstSec;
2655 return first->offset + os->addr - first->addr;
2656 }
2657
assignFileOffsetsBinary()2658 template <class ELFT> void Writer<ELFT>::assignFileOffsetsBinary() {
2659 // Compute the minimum LMA of all non-empty non-NOBITS sections as minAddr.
2660 auto needsOffset = [](OutputSection &sec) {
2661 return sec.type != SHT_NOBITS && (sec.flags & SHF_ALLOC) && sec.size > 0;
2662 };
2663 uint64_t minAddr = UINT64_MAX;
2664 for (OutputSection *sec : outputSections)
2665 if (needsOffset(*sec)) {
2666 sec->offset = sec->getLMA();
2667 minAddr = std::min(minAddr, sec->offset);
2668 }
2669
2670 // Sections are laid out at LMA minus minAddr.
2671 fileSize = 0;
2672 for (OutputSection *sec : outputSections)
2673 if (needsOffset(*sec)) {
2674 sec->offset -= minAddr;
2675 fileSize = std::max(fileSize, sec->offset + sec->size);
2676 }
2677 }
2678
rangeToString(uint64_t addr,uint64_t len)2679 static std::string rangeToString(uint64_t addr, uint64_t len) {
2680 return "[0x" + utohexstr(addr) + ", 0x" + utohexstr(addr + len - 1) + "]";
2681 }
2682
2683 // Assign file offsets to output sections.
assignFileOffsets()2684 template <class ELFT> void Writer<ELFT>::assignFileOffsets() {
2685 Out::programHeaders->offset = Out::elfHeader->size;
2686 uint64_t off = Out::elfHeader->size + Out::programHeaders->size;
2687
2688 PhdrEntry *lastRX = nullptr;
2689 for (Partition &part : partitions)
2690 for (PhdrEntry *p : part.phdrs)
2691 if (p->p_type == PT_LOAD && (p->p_flags & PF_X))
2692 lastRX = p;
2693
2694 // Layout SHF_ALLOC sections before non-SHF_ALLOC sections. A non-SHF_ALLOC
2695 // will not occupy file offsets contained by a PT_LOAD.
2696 for (OutputSection *sec : outputSections) {
2697 if (!(sec->flags & SHF_ALLOC))
2698 continue;
2699 off = computeFileOffset(sec, off);
2700 sec->offset = off;
2701 if (sec->type != SHT_NOBITS)
2702 off += sec->size;
2703
2704 // If this is a last section of the last executable segment and that
2705 // segment is the last loadable segment, align the offset of the
2706 // following section to avoid loading non-segments parts of the file.
2707 if (config->zSeparate != SeparateSegmentKind::None && lastRX &&
2708 lastRX->lastSec == sec)
2709 off = alignToPowerOf2(off, config->maxPageSize);
2710 }
2711 for (OutputSection *osec : outputSections)
2712 if (!(osec->flags & SHF_ALLOC)) {
2713 osec->offset = alignToPowerOf2(off, osec->addralign);
2714 off = osec->offset + osec->size;
2715 }
2716
2717 sectionHeaderOff = alignToPowerOf2(off, config->wordsize);
2718 fileSize = sectionHeaderOff + (outputSections.size() + 1) * sizeof(Elf_Shdr);
2719
2720 // Our logic assumes that sections have rising VA within the same segment.
2721 // With use of linker scripts it is possible to violate this rule and get file
2722 // offset overlaps or overflows. That should never happen with a valid script
2723 // which does not move the location counter backwards and usually scripts do
2724 // not do that. Unfortunately, there are apps in the wild, for example, Linux
2725 // kernel, which control segment distribution explicitly and move the counter
2726 // backwards, so we have to allow doing that to support linking them. We
2727 // perform non-critical checks for overlaps in checkSectionOverlap(), but here
2728 // we want to prevent file size overflows because it would crash the linker.
2729 for (OutputSection *sec : outputSections) {
2730 if (sec->type == SHT_NOBITS)
2731 continue;
2732 if ((sec->offset > fileSize) || (sec->offset + sec->size > fileSize))
2733 error("unable to place section " + sec->name + " at file offset " +
2734 rangeToString(sec->offset, sec->size) +
2735 "; check your linker script for overflows");
2736 }
2737 }
2738
2739 // Finalize the program headers. We call this function after we assign
2740 // file offsets and VAs to all sections.
setPhdrs(Partition & part)2741 template <class ELFT> void Writer<ELFT>::setPhdrs(Partition &part) {
2742 for (PhdrEntry *p : part.phdrs) {
2743 OutputSection *first = p->firstSec;
2744 OutputSection *last = p->lastSec;
2745
2746 // .ARM.exidx sections may not be within a single .ARM.exidx
2747 // output section. We always want to describe just the
2748 // SyntheticSection.
2749 if (part.armExidx && p->p_type == PT_ARM_EXIDX) {
2750 p->p_filesz = part.armExidx->getSize();
2751 p->p_memsz = part.armExidx->getSize();
2752 p->p_offset = first->offset + part.armExidx->outSecOff;
2753 p->p_vaddr = first->addr + part.armExidx->outSecOff;
2754 p->p_align = part.armExidx->addralign;
2755 if (part.elfHeader)
2756 p->p_offset -= part.elfHeader->getParent()->offset;
2757
2758 if (!p->hasLMA)
2759 p->p_paddr = first->getLMA() + part.armExidx->outSecOff;
2760 return;
2761 }
2762
2763 if (first) {
2764 p->p_filesz = last->offset - first->offset;
2765 if (last->type != SHT_NOBITS)
2766 p->p_filesz += last->size;
2767
2768 p->p_memsz = last->addr + last->size - first->addr;
2769 p->p_offset = first->offset;
2770 p->p_vaddr = first->addr;
2771
2772 // File offsets in partitions other than the main partition are relative
2773 // to the offset of the ELF headers. Perform that adjustment now.
2774 if (part.elfHeader)
2775 p->p_offset -= part.elfHeader->getParent()->offset;
2776
2777 if (!p->hasLMA)
2778 p->p_paddr = first->getLMA();
2779 }
2780 }
2781 }
2782
2783 // A helper struct for checkSectionOverlap.
2784 namespace {
2785 struct SectionOffset {
2786 OutputSection *sec;
2787 uint64_t offset;
2788 };
2789 } // namespace
2790
2791 // Check whether sections overlap for a specific address range (file offsets,
2792 // load and virtual addresses).
checkOverlap(StringRef name,std::vector<SectionOffset> & sections,bool isVirtualAddr)2793 static void checkOverlap(StringRef name, std::vector<SectionOffset> §ions,
2794 bool isVirtualAddr) {
2795 llvm::sort(sections, [=](const SectionOffset &a, const SectionOffset &b) {
2796 return a.offset < b.offset;
2797 });
2798
2799 // Finding overlap is easy given a vector is sorted by start position.
2800 // If an element starts before the end of the previous element, they overlap.
2801 for (size_t i = 1, end = sections.size(); i < end; ++i) {
2802 SectionOffset a = sections[i - 1];
2803 SectionOffset b = sections[i];
2804 if (b.offset >= a.offset + a.sec->size)
2805 continue;
2806
2807 // If both sections are in OVERLAY we allow the overlapping of virtual
2808 // addresses, because it is what OVERLAY was designed for.
2809 if (isVirtualAddr && a.sec->inOverlay && b.sec->inOverlay)
2810 continue;
2811
2812 errorOrWarn("section " + a.sec->name + " " + name +
2813 " range overlaps with " + b.sec->name + "\n>>> " + a.sec->name +
2814 " range is " + rangeToString(a.offset, a.sec->size) + "\n>>> " +
2815 b.sec->name + " range is " +
2816 rangeToString(b.offset, b.sec->size));
2817 }
2818 }
2819
2820 // Check for overlapping sections and address overflows.
2821 //
2822 // In this function we check that none of the output sections have overlapping
2823 // file offsets. For SHF_ALLOC sections we also check that the load address
2824 // ranges and the virtual address ranges don't overlap
checkSections()2825 template <class ELFT> void Writer<ELFT>::checkSections() {
2826 // First, check that section's VAs fit in available address space for target.
2827 for (OutputSection *os : outputSections)
2828 if ((os->addr + os->size < os->addr) ||
2829 (!ELFT::Is64Bits && os->addr + os->size > uint64_t(UINT32_MAX) + 1))
2830 errorOrWarn("section " + os->name + " at 0x" + utohexstr(os->addr) +
2831 " of size 0x" + utohexstr(os->size) +
2832 " exceeds available address space");
2833
2834 // Check for overlapping file offsets. In this case we need to skip any
2835 // section marked as SHT_NOBITS. These sections don't actually occupy space in
2836 // the file so Sec->Offset + Sec->Size can overlap with others. If --oformat
2837 // binary is specified only add SHF_ALLOC sections are added to the output
2838 // file so we skip any non-allocated sections in that case.
2839 std::vector<SectionOffset> fileOffs;
2840 for (OutputSection *sec : outputSections)
2841 if (sec->size > 0 && sec->type != SHT_NOBITS &&
2842 (!config->oFormatBinary || (sec->flags & SHF_ALLOC)))
2843 fileOffs.push_back({sec, sec->offset});
2844 checkOverlap("file", fileOffs, false);
2845
2846 // When linking with -r there is no need to check for overlapping virtual/load
2847 // addresses since those addresses will only be assigned when the final
2848 // executable/shared object is created.
2849 if (config->relocatable)
2850 return;
2851
2852 // Checking for overlapping virtual and load addresses only needs to take
2853 // into account SHF_ALLOC sections since others will not be loaded.
2854 // Furthermore, we also need to skip SHF_TLS sections since these will be
2855 // mapped to other addresses at runtime and can therefore have overlapping
2856 // ranges in the file.
2857 std::vector<SectionOffset> vmas;
2858 for (OutputSection *sec : outputSections)
2859 if (sec->size > 0 && (sec->flags & SHF_ALLOC) && !(sec->flags & SHF_TLS))
2860 vmas.push_back({sec, sec->addr});
2861 checkOverlap("virtual address", vmas, true);
2862
2863 // Finally, check that the load addresses don't overlap. This will usually be
2864 // the same as the virtual addresses but can be different when using a linker
2865 // script with AT().
2866 std::vector<SectionOffset> lmas;
2867 for (OutputSection *sec : outputSections)
2868 if (sec->size > 0 && (sec->flags & SHF_ALLOC) && !(sec->flags & SHF_TLS))
2869 lmas.push_back({sec, sec->getLMA()});
2870 checkOverlap("load address", lmas, false);
2871 }
2872
2873 // The entry point address is chosen in the following ways.
2874 //
2875 // 1. the '-e' entry command-line option;
2876 // 2. the ENTRY(symbol) command in a linker control script;
2877 // 3. the value of the symbol _start, if present;
2878 // 4. the number represented by the entry symbol, if it is a number;
2879 // 5. the address 0.
getEntryAddr()2880 static uint64_t getEntryAddr() {
2881 // Case 1, 2 or 3
2882 if (Symbol *b = symtab.find(config->entry))
2883 return b->getVA();
2884
2885 // Case 4
2886 uint64_t addr;
2887 if (to_integer(config->entry, addr))
2888 return addr;
2889
2890 // Case 5
2891 if (config->warnMissingEntry)
2892 warn("cannot find entry symbol " + config->entry +
2893 "; not setting start address");
2894 return 0;
2895 }
2896
getELFType()2897 static uint16_t getELFType() {
2898 if (config->isPic)
2899 return ET_DYN;
2900 if (config->relocatable)
2901 return ET_REL;
2902 return ET_EXEC;
2903 }
2904
writeHeader()2905 template <class ELFT> void Writer<ELFT>::writeHeader() {
2906 writeEhdr<ELFT>(Out::bufferStart, *mainPart);
2907 writePhdrs<ELFT>(Out::bufferStart + sizeof(Elf_Ehdr), *mainPart);
2908
2909 auto *eHdr = reinterpret_cast<Elf_Ehdr *>(Out::bufferStart);
2910 eHdr->e_type = getELFType();
2911 eHdr->e_entry = getEntryAddr();
2912 eHdr->e_shoff = sectionHeaderOff;
2913
2914 // Write the section header table.
2915 //
2916 // The ELF header can only store numbers up to SHN_LORESERVE in the e_shnum
2917 // and e_shstrndx fields. When the value of one of these fields exceeds
2918 // SHN_LORESERVE ELF requires us to put sentinel values in the ELF header and
2919 // use fields in the section header at index 0 to store
2920 // the value. The sentinel values and fields are:
2921 // e_shnum = 0, SHdrs[0].sh_size = number of sections.
2922 // e_shstrndx = SHN_XINDEX, SHdrs[0].sh_link = .shstrtab section index.
2923 auto *sHdrs = reinterpret_cast<Elf_Shdr *>(Out::bufferStart + eHdr->e_shoff);
2924 size_t num = outputSections.size() + 1;
2925 if (num >= SHN_LORESERVE)
2926 sHdrs->sh_size = num;
2927 else
2928 eHdr->e_shnum = num;
2929
2930 uint32_t strTabIndex = in.shStrTab->getParent()->sectionIndex;
2931 if (strTabIndex >= SHN_LORESERVE) {
2932 sHdrs->sh_link = strTabIndex;
2933 eHdr->e_shstrndx = SHN_XINDEX;
2934 } else {
2935 eHdr->e_shstrndx = strTabIndex;
2936 }
2937
2938 for (OutputSection *sec : outputSections)
2939 sec->writeHeaderTo<ELFT>(++sHdrs);
2940 }
2941
2942 // Open a result file.
openFile()2943 template <class ELFT> void Writer<ELFT>::openFile() {
2944 uint64_t maxSize = config->is64 ? INT64_MAX : UINT32_MAX;
2945 if (fileSize != size_t(fileSize) || maxSize < fileSize) {
2946 std::string msg;
2947 raw_string_ostream s(msg);
2948 s << "output file too large: " << Twine(fileSize) << " bytes\n"
2949 << "section sizes:\n";
2950 for (OutputSection *os : outputSections)
2951 s << os->name << ' ' << os->size << "\n";
2952 error(s.str());
2953 return;
2954 }
2955
2956 unlinkAsync(config->outputFile);
2957 unsigned flags = 0;
2958 if (!config->relocatable)
2959 flags |= FileOutputBuffer::F_executable;
2960 if (!config->mmapOutputFile)
2961 flags |= FileOutputBuffer::F_no_mmap;
2962 Expected<std::unique_ptr<FileOutputBuffer>> bufferOrErr =
2963 FileOutputBuffer::create(config->outputFile, fileSize, flags);
2964
2965 if (!bufferOrErr) {
2966 error("failed to open " + config->outputFile + ": " +
2967 llvm::toString(bufferOrErr.takeError()));
2968 return;
2969 }
2970 buffer = std::move(*bufferOrErr);
2971 Out::bufferStart = buffer->getBufferStart();
2972 }
2973
writeSectionsBinary()2974 template <class ELFT> void Writer<ELFT>::writeSectionsBinary() {
2975 parallel::TaskGroup tg;
2976 for (OutputSection *sec : outputSections)
2977 if (sec->flags & SHF_ALLOC)
2978 sec->writeTo<ELFT>(Out::bufferStart + sec->offset, tg);
2979 }
2980
fillTrap(uint8_t * i,uint8_t * end)2981 static void fillTrap(uint8_t *i, uint8_t *end) {
2982 for (; i + 4 <= end; i += 4)
2983 memcpy(i, &target->trapInstr, 4);
2984 }
2985
2986 // Fill the last page of executable segments with trap instructions
2987 // instead of leaving them as zero. Even though it is not required by any
2988 // standard, it is in general a good thing to do for security reasons.
2989 //
2990 // We'll leave other pages in segments as-is because the rest will be
2991 // overwritten by output sections.
writeTrapInstr()2992 template <class ELFT> void Writer<ELFT>::writeTrapInstr() {
2993 for (Partition &part : partitions) {
2994 // Fill the last page.
2995 for (PhdrEntry *p : part.phdrs)
2996 if (p->p_type == PT_LOAD && (p->p_flags & PF_X))
2997 fillTrap(Out::bufferStart +
2998 alignDown(p->firstSec->offset + p->p_filesz, 4),
2999 Out::bufferStart +
3000 alignToPowerOf2(p->firstSec->offset + p->p_filesz,
3001 config->maxPageSize));
3002
3003 // Round up the file size of the last segment to the page boundary iff it is
3004 // an executable segment to ensure that other tools don't accidentally
3005 // trim the instruction padding (e.g. when stripping the file).
3006 PhdrEntry *last = nullptr;
3007 for (PhdrEntry *p : part.phdrs)
3008 if (p->p_type == PT_LOAD)
3009 last = p;
3010
3011 if (last && (last->p_flags & PF_X))
3012 last->p_memsz = last->p_filesz =
3013 alignToPowerOf2(last->p_filesz, config->maxPageSize);
3014 }
3015 }
3016
3017 // Write section contents to a mmap'ed file.
writeSections()3018 template <class ELFT> void Writer<ELFT>::writeSections() {
3019 llvm::TimeTraceScope timeScope("Write sections");
3020
3021 {
3022 // In -r or --emit-relocs mode, write the relocation sections first as in
3023 // ELf_Rel targets we might find out that we need to modify the relocated
3024 // section while doing it.
3025 parallel::TaskGroup tg;
3026 for (OutputSection *sec : outputSections)
3027 if (sec->type == SHT_REL || sec->type == SHT_RELA)
3028 sec->writeTo<ELFT>(Out::bufferStart + sec->offset, tg);
3029 }
3030 {
3031 parallel::TaskGroup tg;
3032 for (OutputSection *sec : outputSections)
3033 if (sec->type != SHT_REL && sec->type != SHT_RELA)
3034 sec->writeTo<ELFT>(Out::bufferStart + sec->offset, tg);
3035 }
3036
3037 // Finally, check that all dynamic relocation addends were written correctly.
3038 if (config->checkDynamicRelocs && config->writeAddends) {
3039 for (OutputSection *sec : outputSections)
3040 if (sec->type == SHT_REL || sec->type == SHT_RELA)
3041 sec->checkDynRelAddends(Out::bufferStart);
3042 }
3043 }
3044
3045 // Computes a hash value of Data using a given hash function.
3046 // In order to utilize multiple cores, we first split data into 1MB
3047 // chunks, compute a hash for each chunk, and then compute a hash value
3048 // of the hash values.
3049 static void
computeHash(llvm::MutableArrayRef<uint8_t> hashBuf,llvm::ArrayRef<uint8_t> data,std::function<void (uint8_t * dest,ArrayRef<uint8_t> arr)> hashFn)3050 computeHash(llvm::MutableArrayRef<uint8_t> hashBuf,
3051 llvm::ArrayRef<uint8_t> data,
3052 std::function<void(uint8_t *dest, ArrayRef<uint8_t> arr)> hashFn) {
3053 std::vector<ArrayRef<uint8_t>> chunks = split(data, 1024 * 1024);
3054 const size_t hashesSize = chunks.size() * hashBuf.size();
3055 std::unique_ptr<uint8_t[]> hashes(new uint8_t[hashesSize]);
3056
3057 // Compute hash values.
3058 parallelFor(0, chunks.size(), [&](size_t i) {
3059 hashFn(hashes.get() + i * hashBuf.size(), chunks[i]);
3060 });
3061
3062 // Write to the final output buffer.
3063 hashFn(hashBuf.data(), ArrayRef(hashes.get(), hashesSize));
3064 }
3065
writeBuildId()3066 template <class ELFT> void Writer<ELFT>::writeBuildId() {
3067 if (!mainPart->buildId || !mainPart->buildId->getParent())
3068 return;
3069
3070 if (config->buildId == BuildIdKind::Hexstring) {
3071 for (Partition &part : partitions)
3072 part.buildId->writeBuildId(config->buildIdVector);
3073 return;
3074 }
3075
3076 // Compute a hash of all sections of the output file.
3077 size_t hashSize = mainPart->buildId->hashSize;
3078 std::unique_ptr<uint8_t[]> buildId(new uint8_t[hashSize]);
3079 MutableArrayRef<uint8_t> output(buildId.get(), hashSize);
3080 llvm::ArrayRef<uint8_t> input{Out::bufferStart, size_t(fileSize)};
3081
3082 // Fedora introduced build ID as "approximation of true uniqueness across all
3083 // binaries that might be used by overlapping sets of people". It does not
3084 // need some security goals that some hash algorithms strive to provide, e.g.
3085 // (second-)preimage and collision resistance. In practice people use 'md5'
3086 // and 'sha1' just for different lengths. Implement them with the more
3087 // efficient BLAKE3.
3088 switch (config->buildId) {
3089 case BuildIdKind::Fast:
3090 computeHash(output, input, [](uint8_t *dest, ArrayRef<uint8_t> arr) {
3091 write64le(dest, xxh3_64bits(arr));
3092 });
3093 break;
3094 case BuildIdKind::Md5:
3095 computeHash(output, input, [&](uint8_t *dest, ArrayRef<uint8_t> arr) {
3096 memcpy(dest, BLAKE3::hash<16>(arr).data(), hashSize);
3097 });
3098 break;
3099 case BuildIdKind::Sha1:
3100 computeHash(output, input, [&](uint8_t *dest, ArrayRef<uint8_t> arr) {
3101 memcpy(dest, BLAKE3::hash<20>(arr).data(), hashSize);
3102 });
3103 break;
3104 case BuildIdKind::Uuid:
3105 if (auto ec = llvm::getRandomBytes(buildId.get(), hashSize))
3106 error("entropy source failure: " + ec.message());
3107 break;
3108 default:
3109 llvm_unreachable("unknown BuildIdKind");
3110 }
3111 for (Partition &part : partitions)
3112 part.buildId->writeBuildId(output);
3113 }
3114
3115 template void elf::createSyntheticSections<ELF32LE>();
3116 template void elf::createSyntheticSections<ELF32BE>();
3117 template void elf::createSyntheticSections<ELF64LE>();
3118 template void elf::createSyntheticSections<ELF64BE>();
3119
3120 template void elf::writeResult<ELF32LE>();
3121 template void elf::writeResult<ELF32BE>();
3122 template void elf::writeResult<ELF64LE>();
3123 template void elf::writeResult<ELF64BE>();
3124