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