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 ; |
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 | } else if (isec->type == SHT_CREL) { |
409 | // The is64=true variant also works with ELF32 since only the r_symidx |
410 | // member is used. |
411 | for (Elf_Crel_Impl<true> r : RelocsCrel<true>(isec->content_)) { |
412 | Symbol &sym = file->getSymbol(symbolIndex: r.r_symidx); |
413 | if (sym.isLocal()) |
414 | sym.used = true; |
415 | } |
416 | } |
417 | } |
418 | } |
419 | } |
420 | |
421 | static bool shouldKeepInSymtab(const Defined &sym) { |
422 | if (sym.isSection()) |
423 | return false; |
424 | |
425 | // If --emit-reloc or -r is given, preserve symbols referenced by relocations |
426 | // from live sections. |
427 | if (sym.used && config->copyRelocs) |
428 | return true; |
429 | |
430 | // Exclude local symbols pointing to .ARM.exidx sections. |
431 | // They are probably mapping symbols "$d", which are optional for these |
432 | // sections. After merging the .ARM.exidx sections, some of these symbols |
433 | // may become dangling. The easiest way to avoid the issue is not to add |
434 | // them to the symbol table from the beginning. |
435 | if (config->emachine == EM_ARM && sym.section && |
436 | sym.section->type == SHT_ARM_EXIDX) |
437 | return false; |
438 | |
439 | if (config->discard == DiscardPolicy::None) |
440 | return true; |
441 | if (config->discard == DiscardPolicy::All) |
442 | return false; |
443 | |
444 | // In ELF assembly .L symbols are normally discarded by the assembler. |
445 | // If the assembler fails to do so, the linker discards them if |
446 | // * --discard-locals is used. |
447 | // * The symbol is in a SHF_MERGE section, which is normally the reason for |
448 | // the assembler keeping the .L symbol. |
449 | if (sym.getName().starts_with(Prefix: ".L" ) && |
450 | (config->discard == DiscardPolicy::Locals || |
451 | (sym.section && (sym.section->flags & SHF_MERGE)))) |
452 | return false; |
453 | return true; |
454 | } |
455 | |
456 | bool lld::elf::includeInSymtab(const Symbol &b) { |
457 | if (auto *d = dyn_cast<Defined>(Val: &b)) { |
458 | // Always include absolute symbols. |
459 | SectionBase *sec = d->section; |
460 | if (!sec) |
461 | return true; |
462 | assert(sec->isLive()); |
463 | |
464 | if (auto *s = dyn_cast<MergeInputSection>(Val: sec)) |
465 | return s->getSectionPiece(offset: d->value).live; |
466 | return true; |
467 | } |
468 | return b.used || !config->gcSections; |
469 | } |
470 | |
471 | // Scan local symbols to: |
472 | // |
473 | // - demote symbols defined relative to /DISCARD/ discarded input sections so |
474 | // that relocations referencing them will lead to errors. |
475 | // - copy eligible symbols to .symTab |
476 | static void demoteAndCopyLocalSymbols() { |
477 | llvm::TimeTraceScope timeScope("Add local symbols" ); |
478 | for (ELFFileBase *file : ctx.objectFiles) { |
479 | DenseMap<SectionBase *, size_t> sectionIndexMap; |
480 | for (Symbol *b : file->getLocalSymbols()) { |
481 | assert(b->isLocal() && "should have been caught in initializeSymbols()" ); |
482 | auto *dr = dyn_cast<Defined>(Val: b); |
483 | if (!dr) |
484 | continue; |
485 | |
486 | if (dr->section && !dr->section->isLive()) |
487 | demoteDefined(sym&: *dr, map&: sectionIndexMap); |
488 | else if (in.symTab && includeInSymtab(b: *b) && shouldKeepInSymtab(sym: *dr)) |
489 | in.symTab->addSymbol(sym: b); |
490 | } |
491 | } |
492 | } |
493 | |
494 | // Create a section symbol for each output section so that we can represent |
495 | // relocations that point to the section. If we know that no relocation is |
496 | // referring to a section (that happens if the section is a synthetic one), we |
497 | // don't create a section symbol for that section. |
498 | template <class ELFT> void Writer<ELFT>::addSectionSymbols() { |
499 | for (SectionCommand *cmd : script->sectionCommands) { |
500 | auto *osd = dyn_cast<OutputDesc>(Val: cmd); |
501 | if (!osd) |
502 | continue; |
503 | OutputSection &osec = osd->osec; |
504 | InputSectionBase *isec = nullptr; |
505 | // Iterate over all input sections and add a STT_SECTION symbol if any input |
506 | // section may be a relocation target. |
507 | for (SectionCommand *cmd : osec.commands) { |
508 | auto *isd = dyn_cast<InputSectionDescription>(Val: cmd); |
509 | if (!isd) |
510 | continue; |
511 | for (InputSectionBase *s : isd->sections) { |
512 | // Relocations are not using REL[A] section symbols. |
513 | if (isStaticRelSecType(type: s->type)) |
514 | continue; |
515 | |
516 | // Unlike other synthetic sections, mergeable output sections contain |
517 | // data copied from input sections, and there may be a relocation |
518 | // pointing to its contents if -r or --emit-reloc is given. |
519 | if (isa<SyntheticSection>(Val: s) && !(s->flags & SHF_MERGE)) |
520 | continue; |
521 | |
522 | isec = s; |
523 | break; |
524 | } |
525 | } |
526 | if (!isec) |
527 | continue; |
528 | |
529 | // Set the symbol to be relative to the output section so that its st_value |
530 | // equals the output section address. Note, there may be a gap between the |
531 | // start of the output section and isec. |
532 | in.symTab->addSymbol(sym: makeDefined(args&: isec->file, args: "" , args: STB_LOCAL, /*stOther=*/args: 0, |
533 | args: STT_SECTION, |
534 | /*value=*/args: 0, /*size=*/args: 0, args: &osec)); |
535 | } |
536 | } |
537 | |
538 | // Today's loaders have a feature to make segments read-only after |
539 | // processing dynamic relocations to enhance security. PT_GNU_RELRO |
540 | // is defined for that. |
541 | // |
542 | // This function returns true if a section needs to be put into a |
543 | // PT_GNU_RELRO segment. |
544 | static bool isRelroSection(const OutputSection *sec) { |
545 | if (!config->zRelro) |
546 | return false; |
547 | if (sec->relro) |
548 | return true; |
549 | |
550 | uint64_t flags = sec->flags; |
551 | |
552 | // Non-allocatable or non-writable sections don't need RELRO because |
553 | // they are not writable or not even mapped to memory in the first place. |
554 | // RELRO is for sections that are essentially read-only but need to |
555 | // be writable only at process startup to allow dynamic linker to |
556 | // apply relocations. |
557 | if (!(flags & SHF_ALLOC) || !(flags & SHF_WRITE)) |
558 | return false; |
559 | |
560 | // Once initialized, TLS data segments are used as data templates |
561 | // for a thread-local storage. For each new thread, runtime |
562 | // allocates memory for a TLS and copy templates there. No thread |
563 | // are supposed to use templates directly. Thus, it can be in RELRO. |
564 | if (flags & SHF_TLS) |
565 | return true; |
566 | |
567 | // .init_array, .preinit_array and .fini_array contain pointers to |
568 | // functions that are executed on process startup or exit. These |
569 | // pointers are set by the static linker, and they are not expected |
570 | // to change at runtime. But if you are an attacker, you could do |
571 | // interesting things by manipulating pointers in .fini_array, for |
572 | // example. So they are put into RELRO. |
573 | uint32_t type = sec->type; |
574 | if (type == SHT_INIT_ARRAY || type == SHT_FINI_ARRAY || |
575 | type == SHT_PREINIT_ARRAY) |
576 | return true; |
577 | |
578 | // .got contains pointers to external symbols. They are resolved by |
579 | // the dynamic linker when a module is loaded into memory, and after |
580 | // that they are not expected to change. So, it can be in RELRO. |
581 | if (in.got && sec == in.got->getParent()) |
582 | return true; |
583 | |
584 | // .toc is a GOT-ish section for PowerPC64. Their contents are accessed |
585 | // through r2 register, which is reserved for that purpose. Since r2 is used |
586 | // for accessing .got as well, .got and .toc need to be close enough in the |
587 | // virtual address space. Usually, .toc comes just after .got. Since we place |
588 | // .got into RELRO, .toc needs to be placed into RELRO too. |
589 | if (sec->name == ".toc" ) |
590 | return true; |
591 | |
592 | // .got.plt contains pointers to external function symbols. They are |
593 | // by default resolved lazily, so we usually cannot put it into RELRO. |
594 | // However, if "-z now" is given, the lazy symbol resolution is |
595 | // disabled, which enables us to put it into RELRO. |
596 | if (sec == in.gotPlt->getParent()) |
597 | return config->zNow; |
598 | |
599 | if (in.relroPadding && sec == in.relroPadding->getParent()) |
600 | return true; |
601 | |
602 | // .dynamic section contains data for the dynamic linker, and |
603 | // there's no need to write to it at runtime, so it's better to put |
604 | // it into RELRO. |
605 | if (sec->name == ".dynamic" ) |
606 | return true; |
607 | |
608 | // Sections with some special names are put into RELRO. This is a |
609 | // bit unfortunate because section names shouldn't be significant in |
610 | // ELF in spirit. But in reality many linker features depend on |
611 | // magic section names. |
612 | StringRef s = sec->name; |
613 | |
614 | bool abiAgnostic = s == ".data.rel.ro" || s == ".bss.rel.ro" || |
615 | s == ".ctors" || s == ".dtors" || s == ".jcr" || |
616 | s == ".eh_frame" || s == ".fini_array" || |
617 | s == ".init_array" || s == ".preinit_array" ; |
618 | |
619 | bool abiSpecific = |
620 | config->osabi == ELFOSABI_OPENBSD && s == ".openbsd.randomdata" ; |
621 | |
622 | return abiAgnostic || abiSpecific; |
623 | } |
624 | |
625 | // We compute a rank for each section. The rank indicates where the |
626 | // section should be placed in the file. Instead of using simple |
627 | // numbers (0,1,2...), we use a series of flags. One for each decision |
628 | // point when placing the section. |
629 | // Using flags has two key properties: |
630 | // * It is easy to check if a give branch was taken. |
631 | // * It is easy two see how similar two ranks are (see getRankProximity). |
632 | enum RankFlags { |
633 | RF_NOT_ADDR_SET = 1 << 27, |
634 | RF_NOT_ALLOC = 1 << 26, |
635 | RF_PARTITION = 1 << 18, // Partition number (8 bits) |
636 | RF_LARGE_ALT = 1 << 15, |
637 | RF_WRITE = 1 << 14, |
638 | RF_EXEC_WRITE = 1 << 13, |
639 | RF_EXEC = 1 << 12, |
640 | RF_RODATA = 1 << 11, |
641 | RF_LARGE = 1 << 10, |
642 | RF_NOT_RELRO = 1 << 9, |
643 | RF_NOT_TLS = 1 << 8, |
644 | RF_BSS = 1 << 7, |
645 | }; |
646 | |
647 | unsigned elf::getSectionRank(OutputSection &osec) { |
648 | unsigned rank = osec.partition * RF_PARTITION; |
649 | |
650 | // We want to put section specified by -T option first, so we |
651 | // can start assigning VA starting from them later. |
652 | if (config->sectionStartMap.count(Key: osec.name)) |
653 | return rank; |
654 | rank |= RF_NOT_ADDR_SET; |
655 | |
656 | // Allocatable sections go first to reduce the total PT_LOAD size and |
657 | // so debug info doesn't change addresses in actual code. |
658 | if (!(osec.flags & SHF_ALLOC)) |
659 | return rank | RF_NOT_ALLOC; |
660 | |
661 | // Sort sections based on their access permission in the following |
662 | // order: R, RX, RXW, RW(RELRO), RW(non-RELRO). |
663 | // |
664 | // Read-only sections come first such that they go in the PT_LOAD covering the |
665 | // program headers at the start of the file. |
666 | // |
667 | // The layout for writable sections is PT_LOAD(PT_GNU_RELRO(.data.rel.ro |
668 | // .bss.rel.ro) | .data .bss), where | marks where page alignment happens. |
669 | // An alternative ordering is PT_LOAD(.data | PT_GNU_RELRO( .data.rel.ro |
670 | // .bss.rel.ro) | .bss), but it may waste more bytes due to 2 alignment |
671 | // places. |
672 | bool isExec = osec.flags & SHF_EXECINSTR; |
673 | bool isWrite = osec.flags & SHF_WRITE; |
674 | |
675 | if (!isWrite && !isExec) { |
676 | // Among PROGBITS sections, place .lrodata further from .text. |
677 | // For -z lrodata-after-bss, place .lrodata after .lbss like GNU ld. This |
678 | // layout has one extra PT_LOAD, but alleviates relocation overflow |
679 | // pressure for absolute relocations referencing small data from -fno-pic |
680 | // relocatable files. |
681 | if (osec.flags & SHF_X86_64_LARGE && config->emachine == EM_X86_64) |
682 | rank |= config->zLrodataAfterBss ? RF_LARGE_ALT : 0; |
683 | else |
684 | rank |= config->zLrodataAfterBss ? 0 : RF_LARGE; |
685 | |
686 | if (osec.type == SHT_LLVM_PART_EHDR) |
687 | ; |
688 | else if (osec.type == SHT_LLVM_PART_PHDR) |
689 | rank |= 1; |
690 | else if (osec.name == ".interp" ) |
691 | rank |= 2; |
692 | // Put .note sections at the beginning so that they are likely to be |
693 | // included in a truncate core file. In particular, .note.gnu.build-id, if |
694 | // available, can identify the object file. |
695 | else if (osec.type == SHT_NOTE) |
696 | rank |= 3; |
697 | // Make PROGBITS sections (e.g .rodata .eh_frame) closer to .text to |
698 | // alleviate relocation overflow pressure. Large special sections such as |
699 | // .dynstr and .dynsym can be away from .text. |
700 | else if (osec.type != SHT_PROGBITS) |
701 | rank |= 4; |
702 | else |
703 | rank |= RF_RODATA; |
704 | } else if (isExec) { |
705 | rank |= isWrite ? RF_EXEC_WRITE : RF_EXEC; |
706 | } else { |
707 | rank |= RF_WRITE; |
708 | // The TLS initialization block needs to be a single contiguous block. Place |
709 | // TLS sections directly before the other RELRO sections. |
710 | if (!(osec.flags & SHF_TLS)) |
711 | rank |= RF_NOT_TLS; |
712 | if (isRelroSection(sec: &osec)) |
713 | osec.relro = true; |
714 | else |
715 | rank |= RF_NOT_RELRO; |
716 | // Place .ldata and .lbss after .bss. Making .bss closer to .text |
717 | // alleviates relocation overflow pressure. |
718 | // For -z lrodata-after-bss, place .lbss/.lrodata/.ldata after .bss. |
719 | // .bss/.lbss being adjacent reuses the NOBITS size optimization. |
720 | if (osec.flags & SHF_X86_64_LARGE && config->emachine == EM_X86_64) { |
721 | rank |= config->zLrodataAfterBss |
722 | ? (osec.type == SHT_NOBITS ? 1 : RF_LARGE_ALT) |
723 | : RF_LARGE; |
724 | } |
725 | } |
726 | |
727 | // Within TLS sections, or within other RelRo sections, or within non-RelRo |
728 | // sections, place non-NOBITS sections first. |
729 | if (osec.type == SHT_NOBITS) |
730 | rank |= RF_BSS; |
731 | |
732 | // Some architectures have additional ordering restrictions for sections |
733 | // within the same PT_LOAD. |
734 | if (config->emachine == EM_PPC64) { |
735 | // PPC64 has a number of special SHT_PROGBITS+SHF_ALLOC+SHF_WRITE sections |
736 | // that we would like to make sure appear is a specific order to maximize |
737 | // their coverage by a single signed 16-bit offset from the TOC base |
738 | // pointer. |
739 | StringRef name = osec.name; |
740 | if (name == ".got" ) |
741 | rank |= 1; |
742 | else if (name == ".toc" ) |
743 | rank |= 2; |
744 | } |
745 | |
746 | if (config->emachine == EM_MIPS) { |
747 | if (osec.name != ".got" ) |
748 | rank |= 1; |
749 | // All sections with SHF_MIPS_GPREL flag should be grouped together |
750 | // because data in these sections is addressable with a gp relative address. |
751 | if (osec.flags & SHF_MIPS_GPREL) |
752 | rank |= 2; |
753 | } |
754 | |
755 | if (config->emachine == EM_RISCV) { |
756 | // .sdata and .sbss are placed closer to make GP relaxation more profitable |
757 | // and match GNU ld. |
758 | StringRef name = osec.name; |
759 | if (name == ".sdata" || (osec.type == SHT_NOBITS && name != ".sbss" )) |
760 | rank |= 1; |
761 | } |
762 | |
763 | return rank; |
764 | } |
765 | |
766 | static bool compareSections(const SectionCommand *aCmd, |
767 | const SectionCommand *bCmd) { |
768 | const OutputSection *a = &cast<OutputDesc>(Val: aCmd)->osec; |
769 | const OutputSection *b = &cast<OutputDesc>(Val: bCmd)->osec; |
770 | |
771 | if (a->sortRank != b->sortRank) |
772 | return a->sortRank < b->sortRank; |
773 | |
774 | if (!(a->sortRank & RF_NOT_ADDR_SET)) |
775 | return config->sectionStartMap.lookup(Key: a->name) < |
776 | config->sectionStartMap.lookup(Key: b->name); |
777 | return false; |
778 | } |
779 | |
780 | void PhdrEntry::add(OutputSection *sec) { |
781 | lastSec = sec; |
782 | if (!firstSec) |
783 | firstSec = sec; |
784 | p_align = std::max(a: p_align, b: sec->addralign); |
785 | if (p_type == PT_LOAD) |
786 | sec->ptLoad = this; |
787 | } |
788 | |
789 | // A statically linked position-dependent executable should only contain |
790 | // IRELATIVE relocations and no other dynamic relocations. Encapsulation symbols |
791 | // __rel[a]_iplt_{start,end} will be defined for .rel[a].dyn, to be |
792 | // processed by the libc runtime. Other executables or DSOs use dynamic tags |
793 | // instead. |
794 | template <class ELFT> void Writer<ELFT>::addRelIpltSymbols() { |
795 | if (config->isPic) |
796 | return; |
797 | |
798 | // __rela_iplt_{start,end} are initially defined relative to dummy section 0. |
799 | // We'll override Out::elfHeader with relaDyn later when we are sure that |
800 | // .rela.dyn will be present in the output. |
801 | std::string name = config->isRela ? "__rela_iplt_start" : "__rel_iplt_start" ; |
802 | ElfSym::relaIpltStart = |
803 | addOptionalRegular(name, sec: Out::elfHeader, val: 0, stOther: STV_HIDDEN); |
804 | name.replace(pos: name.size() - 5, n1: 5, s: "end" ); |
805 | ElfSym::relaIpltEnd = addOptionalRegular(name, sec: Out::elfHeader, val: 0, stOther: STV_HIDDEN); |
806 | } |
807 | |
808 | // This function generates assignments for predefined symbols (e.g. _end or |
809 | // _etext) and inserts them into the commands sequence to be processed at the |
810 | // appropriate time. This ensures that the value is going to be correct by the |
811 | // time any references to these symbols are processed and is equivalent to |
812 | // defining these symbols explicitly in the linker script. |
813 | template <class ELFT> void Writer<ELFT>::setReservedSymbolSections() { |
814 | if (ElfSym::globalOffsetTable) { |
815 | // The _GLOBAL_OFFSET_TABLE_ symbol is defined by target convention usually |
816 | // to the start of the .got or .got.plt section. |
817 | InputSection *sec = in.gotPlt.get(); |
818 | if (!target->gotBaseSymInGotPlt) |
819 | sec = in.mipsGot ? cast<InputSection>(Val: in.mipsGot.get()) |
820 | : cast<InputSection>(Val: in.got.get()); |
821 | ElfSym::globalOffsetTable->section = sec; |
822 | } |
823 | |
824 | // .rela_iplt_{start,end} mark the start and the end of .rel[a].dyn. |
825 | if (ElfSym::relaIpltStart && mainPart->relaDyn->isNeeded()) { |
826 | ElfSym::relaIpltStart->section = mainPart->relaDyn.get(); |
827 | ElfSym::relaIpltEnd->section = mainPart->relaDyn.get(); |
828 | ElfSym::relaIpltEnd->value = mainPart->relaDyn->getSize(); |
829 | } |
830 | |
831 | PhdrEntry *last = nullptr; |
832 | OutputSection *lastRO = nullptr; |
833 | auto isLarge = [](OutputSection *osec) { |
834 | return config->emachine == EM_X86_64 && osec->flags & SHF_X86_64_LARGE; |
835 | }; |
836 | for (Partition &part : partitions) { |
837 | for (PhdrEntry *p : part.phdrs) { |
838 | if (p->p_type != PT_LOAD) |
839 | continue; |
840 | last = p; |
841 | if (!(p->p_flags & PF_W) && p->lastSec && !isLarge(p->lastSec)) |
842 | lastRO = p->lastSec; |
843 | } |
844 | } |
845 | |
846 | if (lastRO) { |
847 | // _etext is the first location after the last read-only loadable segment |
848 | // that does not contain large sections. |
849 | if (ElfSym::etext1) |
850 | ElfSym::etext1->section = lastRO; |
851 | if (ElfSym::etext2) |
852 | ElfSym::etext2->section = lastRO; |
853 | } |
854 | |
855 | if (last) { |
856 | // _edata points to the end of the last non-large mapped initialized |
857 | // section. |
858 | OutputSection *edata = nullptr; |
859 | for (OutputSection *os : outputSections) { |
860 | if (os->type != SHT_NOBITS && !isLarge(os)) |
861 | edata = os; |
862 | if (os == last->lastSec) |
863 | break; |
864 | } |
865 | |
866 | if (ElfSym::edata1) |
867 | ElfSym::edata1->section = edata; |
868 | if (ElfSym::edata2) |
869 | ElfSym::edata2->section = edata; |
870 | |
871 | // _end is the first location after the uninitialized data region. |
872 | if (ElfSym::end1) |
873 | ElfSym::end1->section = last->lastSec; |
874 | if (ElfSym::end2) |
875 | ElfSym::end2->section = last->lastSec; |
876 | } |
877 | |
878 | if (ElfSym::bss) { |
879 | // On RISC-V, set __bss_start to the start of .sbss if present. |
880 | OutputSection *sbss = |
881 | config->emachine == EM_RISCV ? findSection(name: ".sbss" ) : nullptr; |
882 | ElfSym::bss->section = sbss ? sbss : findSection(name: ".bss" ); |
883 | } |
884 | |
885 | // Setup MIPS _gp_disp/__gnu_local_gp symbols which should |
886 | // be equal to the _gp symbol's value. |
887 | if (ElfSym::mipsGp) { |
888 | // Find GP-relative section with the lowest address |
889 | // and use this address to calculate default _gp value. |
890 | for (OutputSection *os : outputSections) { |
891 | if (os->flags & SHF_MIPS_GPREL) { |
892 | ElfSym::mipsGp->section = os; |
893 | ElfSym::mipsGp->value = 0x7ff0; |
894 | break; |
895 | } |
896 | } |
897 | } |
898 | } |
899 | |
900 | // We want to find how similar two ranks are. |
901 | // The more branches in getSectionRank that match, the more similar they are. |
902 | // Since each branch corresponds to a bit flag, we can just use |
903 | // countLeadingZeros. |
904 | static int getRankProximity(OutputSection *a, SectionCommand *b) { |
905 | auto *osd = dyn_cast<OutputDesc>(Val: b); |
906 | return (osd && osd->osec.hasInputSections) |
907 | ? llvm::countl_zero(Val: a->sortRank ^ osd->osec.sortRank) |
908 | : -1; |
909 | } |
910 | |
911 | // When placing orphan sections, we want to place them after symbol assignments |
912 | // so that an orphan after |
913 | // begin_foo = .; |
914 | // foo : { *(foo) } |
915 | // end_foo = .; |
916 | // doesn't break the intended meaning of the begin/end symbols. |
917 | // We don't want to go over sections since findOrphanPos is the |
918 | // one in charge of deciding the order of the sections. |
919 | // We don't want to go over changes to '.', since doing so in |
920 | // rx_sec : { *(rx_sec) } |
921 | // . = ALIGN(0x1000); |
922 | // /* The RW PT_LOAD starts here*/ |
923 | // rw_sec : { *(rw_sec) } |
924 | // would mean that the RW PT_LOAD would become unaligned. |
925 | static bool shouldSkip(SectionCommand *cmd) { |
926 | if (auto *assign = dyn_cast<SymbolAssignment>(Val: cmd)) |
927 | return assign->name != "." ; |
928 | return false; |
929 | } |
930 | |
931 | // We want to place orphan sections so that they share as much |
932 | // characteristics with their neighbors as possible. For example, if |
933 | // both are rw, or both are tls. |
934 | static SmallVectorImpl<SectionCommand *>::iterator |
935 | findOrphanPos(SmallVectorImpl<SectionCommand *>::iterator b, |
936 | SmallVectorImpl<SectionCommand *>::iterator e) { |
937 | // Place non-alloc orphan sections at the end. This matches how we assign file |
938 | // offsets to non-alloc sections. |
939 | OutputSection *sec = &cast<OutputDesc>(Val: *e)->osec; |
940 | if (!(sec->flags & SHF_ALLOC)) |
941 | return e; |
942 | |
943 | // As a special case, place .relro_padding before the SymbolAssignment using |
944 | // DATA_SEGMENT_RELRO_END, if present. |
945 | if (in.relroPadding && sec == in.relroPadding->getParent()) { |
946 | auto i = std::find_if(first: b, last: e, pred: [=](SectionCommand *a) { |
947 | if (auto *assign = dyn_cast<SymbolAssignment>(Val: a)) |
948 | return assign->dataSegmentRelroEnd; |
949 | return false; |
950 | }); |
951 | if (i != e) |
952 | return i; |
953 | } |
954 | |
955 | // Find the most similar output section as the anchor. Rank Proximity is a |
956 | // value in the range [-1, 32] where [0, 32] indicates potential anchors (0: |
957 | // least similar; 32: identical). -1 means not an anchor. |
958 | // |
959 | // In the event of proximity ties, we select the first or last section |
960 | // depending on whether the orphan's rank is smaller. |
961 | int maxP = 0; |
962 | auto i = e; |
963 | for (auto j = b; j != e; ++j) { |
964 | int p = getRankProximity(a: sec, b: *j); |
965 | if (p > maxP || |
966 | (p == maxP && cast<OutputDesc>(Val: *j)->osec.sortRank <= sec->sortRank)) { |
967 | maxP = p; |
968 | i = j; |
969 | } |
970 | } |
971 | if (i == e) |
972 | return e; |
973 | |
974 | auto isOutputSecWithInputSections = [](SectionCommand *cmd) { |
975 | auto *osd = dyn_cast<OutputDesc>(Val: cmd); |
976 | return osd && osd->osec.hasInputSections; |
977 | }; |
978 | |
979 | // Then, scan backward or forward through the script for a suitable insertion |
980 | // point. If i's rank is larger, the orphan section can be placed before i. |
981 | // |
982 | // However, don't do this if custom program headers are defined. Otherwise, |
983 | // adding the orphan to a previous segment can change its flags, for example, |
984 | // making a read-only segment writable. If memory regions are defined, an |
985 | // orphan section should continue the same region as the found section to |
986 | // better resemble the behavior of GNU ld. |
987 | bool mustAfter = script->hasPhdrsCommands() || !script->memoryRegions.empty(); |
988 | if (cast<OutputDesc>(Val: *i)->osec.sortRank <= sec->sortRank || mustAfter) { |
989 | for (auto j = ++i; j != e; ++j) { |
990 | if (!isOutputSecWithInputSections(*j)) |
991 | continue; |
992 | if (getRankProximity(a: sec, b: *j) != maxP) |
993 | break; |
994 | i = j + 1; |
995 | } |
996 | } else { |
997 | for (; i != b; --i) |
998 | if (isOutputSecWithInputSections(i[-1])) |
999 | break; |
1000 | } |
1001 | |
1002 | // As a special case, if the orphan section is the last section, put |
1003 | // it at the very end, past any other commands. |
1004 | // This matches bfd's behavior and is convenient when the linker script fully |
1005 | // specifies the start of the file, but doesn't care about the end (the non |
1006 | // alloc sections for example). |
1007 | if (std::find_if(first: i, last: e, pred: isOutputSecWithInputSections) == e) |
1008 | return e; |
1009 | |
1010 | while (i != e && shouldSkip(cmd: *i)) |
1011 | ++i; |
1012 | return i; |
1013 | } |
1014 | |
1015 | // Adds random priorities to sections not already in the map. |
1016 | static void maybeShuffle(DenseMap<const InputSectionBase *, int> &order) { |
1017 | if (config->shuffleSections.empty()) |
1018 | return; |
1019 | |
1020 | SmallVector<InputSectionBase *, 0> matched, sections = ctx.inputSections; |
1021 | matched.reserve(N: sections.size()); |
1022 | for (const auto &patAndSeed : config->shuffleSections) { |
1023 | matched.clear(); |
1024 | for (InputSectionBase *sec : sections) |
1025 | if (patAndSeed.first.match(S: sec->name)) |
1026 | matched.push_back(Elt: sec); |
1027 | const uint32_t seed = patAndSeed.second; |
1028 | if (seed == UINT32_MAX) { |
1029 | // If --shuffle-sections <section-glob>=-1, reverse the section order. The |
1030 | // section order is stable even if the number of sections changes. This is |
1031 | // useful to catch issues like static initialization order fiasco |
1032 | // reliably. |
1033 | std::reverse(first: matched.begin(), last: matched.end()); |
1034 | } else { |
1035 | std::mt19937 g(seed ? seed : std::random_device()()); |
1036 | llvm::shuffle(first: matched.begin(), last: matched.end(), g); |
1037 | } |
1038 | size_t i = 0; |
1039 | for (InputSectionBase *&sec : sections) |
1040 | if (patAndSeed.first.match(S: sec->name)) |
1041 | sec = matched[i++]; |
1042 | } |
1043 | |
1044 | // Existing priorities are < 0, so use priorities >= 0 for the missing |
1045 | // sections. |
1046 | int prio = 0; |
1047 | for (InputSectionBase *sec : sections) { |
1048 | if (order.try_emplace(Key: sec, Args&: prio).second) |
1049 | ++prio; |
1050 | } |
1051 | } |
1052 | |
1053 | // Builds section order for handling --symbol-ordering-file. |
1054 | static DenseMap<const InputSectionBase *, int> buildSectionOrder() { |
1055 | DenseMap<const InputSectionBase *, int> sectionOrder; |
1056 | // Use the rarely used option --call-graph-ordering-file to sort sections. |
1057 | if (!config->callGraphProfile.empty()) |
1058 | return computeCallGraphProfileOrder(); |
1059 | |
1060 | if (config->symbolOrderingFile.empty()) |
1061 | return sectionOrder; |
1062 | |
1063 | struct SymbolOrderEntry { |
1064 | int priority; |
1065 | bool present; |
1066 | }; |
1067 | |
1068 | // Build a map from symbols to their priorities. Symbols that didn't |
1069 | // appear in the symbol ordering file have the lowest priority 0. |
1070 | // All explicitly mentioned symbols have negative (higher) priorities. |
1071 | DenseMap<CachedHashStringRef, SymbolOrderEntry> symbolOrder; |
1072 | int priority = -config->symbolOrderingFile.size(); |
1073 | for (StringRef s : config->symbolOrderingFile) |
1074 | symbolOrder.insert(KV: {CachedHashStringRef(s), {.priority: priority++, .present: false}}); |
1075 | |
1076 | // Build a map from sections to their priorities. |
1077 | auto addSym = [&](Symbol &sym) { |
1078 | auto it = symbolOrder.find(Val: CachedHashStringRef(sym.getName())); |
1079 | if (it == symbolOrder.end()) |
1080 | return; |
1081 | SymbolOrderEntry &ent = it->second; |
1082 | ent.present = true; |
1083 | |
1084 | maybeWarnUnorderableSymbol(sym: &sym); |
1085 | |
1086 | if (auto *d = dyn_cast<Defined>(Val: &sym)) { |
1087 | if (auto *sec = dyn_cast_or_null<InputSectionBase>(Val: d->section)) { |
1088 | int &priority = sectionOrder[cast<InputSectionBase>(Val: sec)]; |
1089 | priority = std::min(a: priority, b: ent.priority); |
1090 | } |
1091 | } |
1092 | }; |
1093 | |
1094 | // We want both global and local symbols. We get the global ones from the |
1095 | // symbol table and iterate the object files for the local ones. |
1096 | for (Symbol *sym : symtab.getSymbols()) |
1097 | addSym(*sym); |
1098 | |
1099 | for (ELFFileBase *file : ctx.objectFiles) |
1100 | for (Symbol *sym : file->getLocalSymbols()) |
1101 | addSym(*sym); |
1102 | |
1103 | if (config->warnSymbolOrdering) |
1104 | for (auto orderEntry : symbolOrder) |
1105 | if (!orderEntry.second.present) |
1106 | warn(msg: "symbol ordering file: no such symbol: " + orderEntry.first.val()); |
1107 | |
1108 | return sectionOrder; |
1109 | } |
1110 | |
1111 | // Sorts the sections in ISD according to the provided section order. |
1112 | static void |
1113 | sortISDBySectionOrder(InputSectionDescription *isd, |
1114 | const DenseMap<const InputSectionBase *, int> &order, |
1115 | bool executableOutputSection) { |
1116 | SmallVector<InputSection *, 0> unorderedSections; |
1117 | SmallVector<std::pair<InputSection *, int>, 0> orderedSections; |
1118 | uint64_t unorderedSize = 0; |
1119 | uint64_t totalSize = 0; |
1120 | |
1121 | for (InputSection *isec : isd->sections) { |
1122 | if (executableOutputSection) |
1123 | totalSize += isec->getSize(); |
1124 | auto i = order.find(Val: isec); |
1125 | if (i == order.end()) { |
1126 | unorderedSections.push_back(Elt: isec); |
1127 | unorderedSize += isec->getSize(); |
1128 | continue; |
1129 | } |
1130 | orderedSections.push_back(Elt: {isec, i->second}); |
1131 | } |
1132 | llvm::sort(C&: orderedSections, Comp: llvm::less_second()); |
1133 | |
1134 | // Find an insertion point for the ordered section list in the unordered |
1135 | // section list. On targets with limited-range branches, this is the mid-point |
1136 | // of the unordered section list. This decreases the likelihood that a range |
1137 | // extension thunk will be needed to enter or exit the ordered region. If the |
1138 | // ordered section list is a list of hot functions, we can generally expect |
1139 | // the ordered functions to be called more often than the unordered functions, |
1140 | // making it more likely that any particular call will be within range, and |
1141 | // therefore reducing the number of thunks required. |
1142 | // |
1143 | // For example, imagine that you have 8MB of hot code and 32MB of cold code. |
1144 | // If the layout is: |
1145 | // |
1146 | // 8MB hot |
1147 | // 32MB cold |
1148 | // |
1149 | // only the first 8-16MB of the cold code (depending on which hot function it |
1150 | // is actually calling) can call the hot code without a range extension thunk. |
1151 | // However, if we use this layout: |
1152 | // |
1153 | // 16MB cold |
1154 | // 8MB hot |
1155 | // 16MB cold |
1156 | // |
1157 | // both the last 8-16MB of the first block of cold code and the first 8-16MB |
1158 | // of the second block of cold code can call the hot code without a thunk. So |
1159 | // we effectively double the amount of code that could potentially call into |
1160 | // the hot code without a thunk. |
1161 | // |
1162 | // The above is not necessary if total size of input sections in this "isd" |
1163 | // is small. Note that we assume all input sections are executable if the |
1164 | // output section is executable (which is not always true but supposed to |
1165 | // cover most cases). |
1166 | size_t insPt = 0; |
1167 | if (executableOutputSection && !orderedSections.empty() && |
1168 | target->getThunkSectionSpacing() && |
1169 | totalSize >= target->getThunkSectionSpacing()) { |
1170 | uint64_t unorderedPos = 0; |
1171 | for (; insPt != unorderedSections.size(); ++insPt) { |
1172 | unorderedPos += unorderedSections[insPt]->getSize(); |
1173 | if (unorderedPos > unorderedSize / 2) |
1174 | break; |
1175 | } |
1176 | } |
1177 | |
1178 | isd->sections.clear(); |
1179 | for (InputSection *isec : ArrayRef(unorderedSections).slice(N: 0, M: insPt)) |
1180 | isd->sections.push_back(Elt: isec); |
1181 | for (std::pair<InputSection *, int> p : orderedSections) |
1182 | isd->sections.push_back(Elt: p.first); |
1183 | for (InputSection *isec : ArrayRef(unorderedSections).slice(N: insPt)) |
1184 | isd->sections.push_back(Elt: isec); |
1185 | } |
1186 | |
1187 | static void sortSection(OutputSection &osec, |
1188 | const DenseMap<const InputSectionBase *, int> &order) { |
1189 | StringRef name = osec.name; |
1190 | |
1191 | // Never sort these. |
1192 | if (name == ".init" || name == ".fini" ) |
1193 | return; |
1194 | |
1195 | // Sort input sections by priority using the list provided by |
1196 | // --symbol-ordering-file or --shuffle-sections=. This is a least significant |
1197 | // digit radix sort. The sections may be sorted stably again by a more |
1198 | // significant key. |
1199 | if (!order.empty()) |
1200 | for (SectionCommand *b : osec.commands) |
1201 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: b)) |
1202 | sortISDBySectionOrder(isd, order, executableOutputSection: osec.flags & SHF_EXECINSTR); |
1203 | |
1204 | if (script->hasSectionsCommand) |
1205 | return; |
1206 | |
1207 | if (name == ".init_array" || name == ".fini_array" ) { |
1208 | osec.sortInitFini(); |
1209 | } else if (name == ".ctors" || name == ".dtors" ) { |
1210 | osec.sortCtorsDtors(); |
1211 | } else if (config->emachine == EM_PPC64 && name == ".toc" ) { |
1212 | // .toc is allocated just after .got and is accessed using GOT-relative |
1213 | // relocations. Object files compiled with small code model have an |
1214 | // addressable range of [.got, .got + 0xFFFC] for GOT-relative relocations. |
1215 | // To reduce the risk of relocation overflow, .toc contents are sorted so |
1216 | // that sections having smaller relocation offsets are at beginning of .toc |
1217 | assert(osec.commands.size() == 1); |
1218 | auto *isd = cast<InputSectionDescription>(Val: osec.commands[0]); |
1219 | llvm::stable_sort(Range&: isd->sections, |
1220 | C: [](const InputSection *a, const InputSection *b) -> bool { |
1221 | return a->file->ppc64SmallCodeModelTocRelocs && |
1222 | !b->file->ppc64SmallCodeModelTocRelocs; |
1223 | }); |
1224 | } |
1225 | } |
1226 | |
1227 | // If no layout was provided by linker script, we want to apply default |
1228 | // sorting for special input sections. This also handles --symbol-ordering-file. |
1229 | template <class ELFT> void Writer<ELFT>::sortInputSections() { |
1230 | // Build the order once since it is expensive. |
1231 | DenseMap<const InputSectionBase *, int> order = buildSectionOrder(); |
1232 | maybeShuffle(order); |
1233 | for (SectionCommand *cmd : script->sectionCommands) |
1234 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) |
1235 | sortSection(osec&: osd->osec, order); |
1236 | } |
1237 | |
1238 | template <class ELFT> void Writer<ELFT>::sortSections() { |
1239 | llvm::TimeTraceScope timeScope("Sort sections" ); |
1240 | |
1241 | // Don't sort if using -r. It is not necessary and we want to preserve the |
1242 | // relative order for SHF_LINK_ORDER sections. |
1243 | if (config->relocatable) { |
1244 | script->adjustOutputSections(); |
1245 | return; |
1246 | } |
1247 | |
1248 | sortInputSections(); |
1249 | |
1250 | for (SectionCommand *cmd : script->sectionCommands) |
1251 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) |
1252 | osd->osec.sortRank = getSectionRank(osec&: osd->osec); |
1253 | if (!script->hasSectionsCommand) { |
1254 | // OutputDescs are mostly contiguous, but may be interleaved with |
1255 | // SymbolAssignments in the presence of INSERT commands. |
1256 | auto mid = std::stable_partition( |
1257 | script->sectionCommands.begin(), script->sectionCommands.end(), |
1258 | [](SectionCommand *cmd) { return isa<OutputDesc>(Val: cmd); }); |
1259 | std::stable_sort(script->sectionCommands.begin(), mid, compareSections); |
1260 | } |
1261 | |
1262 | // Process INSERT commands and update output section attributes. From this |
1263 | // point onwards the order of script->sectionCommands is fixed. |
1264 | script->processInsertCommands(); |
1265 | script->adjustOutputSections(); |
1266 | |
1267 | if (script->hasSectionsCommand) |
1268 | sortOrphanSections(); |
1269 | |
1270 | script->adjustSectionsAfterSorting(); |
1271 | } |
1272 | |
1273 | template <class ELFT> void Writer<ELFT>::sortOrphanSections() { |
1274 | // Orphan sections are sections present in the input files which are |
1275 | // not explicitly placed into the output file by the linker script. |
1276 | // |
1277 | // The sections in the linker script are already in the correct |
1278 | // order. We have to figuere out where to insert the orphan |
1279 | // sections. |
1280 | // |
1281 | // The order of the sections in the script is arbitrary and may not agree with |
1282 | // compareSections. This means that we cannot easily define a strict weak |
1283 | // ordering. To see why, consider a comparison of a section in the script and |
1284 | // one not in the script. We have a two simple options: |
1285 | // * Make them equivalent (a is not less than b, and b is not less than a). |
1286 | // The problem is then that equivalence has to be transitive and we can |
1287 | // have sections a, b and c with only b in a script and a less than c |
1288 | // which breaks this property. |
1289 | // * Use compareSectionsNonScript. Given that the script order doesn't have |
1290 | // to match, we can end up with sections a, b, c, d where b and c are in the |
1291 | // script and c is compareSectionsNonScript less than b. In which case d |
1292 | // can be equivalent to c, a to b and d < a. As a concrete example: |
1293 | // .a (rx) # not in script |
1294 | // .b (rx) # in script |
1295 | // .c (ro) # in script |
1296 | // .d (ro) # not in script |
1297 | // |
1298 | // The way we define an order then is: |
1299 | // * Sort only the orphan sections. They are in the end right now. |
1300 | // * Move each orphan section to its preferred position. We try |
1301 | // to put each section in the last position where it can share |
1302 | // a PT_LOAD. |
1303 | // |
1304 | // There is some ambiguity as to where exactly a new entry should be |
1305 | // inserted, because Commands contains not only output section |
1306 | // commands but also other types of commands such as symbol assignment |
1307 | // expressions. There's no correct answer here due to the lack of the |
1308 | // formal specification of the linker script. We use heuristics to |
1309 | // determine whether a new output command should be added before or |
1310 | // after another commands. For the details, look at shouldSkip |
1311 | // function. |
1312 | |
1313 | auto i = script->sectionCommands.begin(); |
1314 | auto e = script->sectionCommands.end(); |
1315 | auto nonScriptI = std::find_if(i, e, [](SectionCommand *cmd) { |
1316 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) |
1317 | return osd->osec.sectionIndex == UINT32_MAX; |
1318 | return false; |
1319 | }); |
1320 | |
1321 | // Sort the orphan sections. |
1322 | std::stable_sort(nonScriptI, e, compareSections); |
1323 | |
1324 | // As a horrible special case, skip the first . assignment if it is before any |
1325 | // section. We do this because it is common to set a load address by starting |
1326 | // the script with ". = 0xabcd" and the expectation is that every section is |
1327 | // after that. |
1328 | auto firstSectionOrDotAssignment = |
1329 | std::find_if(i, e, [](SectionCommand *cmd) { return !shouldSkip(cmd); }); |
1330 | if (firstSectionOrDotAssignment != e && |
1331 | isa<SymbolAssignment>(**firstSectionOrDotAssignment)) |
1332 | ++firstSectionOrDotAssignment; |
1333 | i = firstSectionOrDotAssignment; |
1334 | |
1335 | while (nonScriptI != e) { |
1336 | auto pos = findOrphanPos(i, nonScriptI); |
1337 | OutputSection *orphan = &cast<OutputDesc>(*nonScriptI)->osec; |
1338 | |
1339 | // As an optimization, find all sections with the same sort rank |
1340 | // and insert them with one rotate. |
1341 | unsigned rank = orphan->sortRank; |
1342 | auto end = std::find_if(nonScriptI + 1, e, [=](SectionCommand *cmd) { |
1343 | return cast<OutputDesc>(Val: cmd)->osec.sortRank != rank; |
1344 | }); |
1345 | std::rotate(pos, nonScriptI, end); |
1346 | nonScriptI = end; |
1347 | } |
1348 | } |
1349 | |
1350 | static bool compareByFilePosition(InputSection *a, InputSection *b) { |
1351 | InputSection *la = a->flags & SHF_LINK_ORDER ? a->getLinkOrderDep() : nullptr; |
1352 | InputSection *lb = b->flags & SHF_LINK_ORDER ? b->getLinkOrderDep() : nullptr; |
1353 | // SHF_LINK_ORDER sections with non-zero sh_link are ordered before |
1354 | // non-SHF_LINK_ORDER sections and SHF_LINK_ORDER sections with zero sh_link. |
1355 | if (!la || !lb) |
1356 | return la && !lb; |
1357 | OutputSection *aOut = la->getParent(); |
1358 | OutputSection *bOut = lb->getParent(); |
1359 | |
1360 | if (aOut == bOut) |
1361 | return la->outSecOff < lb->outSecOff; |
1362 | if (aOut->addr == bOut->addr) |
1363 | return aOut->sectionIndex < bOut->sectionIndex; |
1364 | return aOut->addr < bOut->addr; |
1365 | } |
1366 | |
1367 | template <class ELFT> void Writer<ELFT>::resolveShfLinkOrder() { |
1368 | llvm::TimeTraceScope timeScope("Resolve SHF_LINK_ORDER" ); |
1369 | for (OutputSection *sec : outputSections) { |
1370 | if (!(sec->flags & SHF_LINK_ORDER)) |
1371 | continue; |
1372 | |
1373 | // The ARM.exidx section use SHF_LINK_ORDER, but we have consolidated |
1374 | // this processing inside the ARMExidxsyntheticsection::finalizeContents(). |
1375 | if (!config->relocatable && config->emachine == EM_ARM && |
1376 | sec->type == SHT_ARM_EXIDX) |
1377 | continue; |
1378 | |
1379 | // Link order may be distributed across several InputSectionDescriptions. |
1380 | // Sorting is performed separately. |
1381 | SmallVector<InputSection **, 0> scriptSections; |
1382 | SmallVector<InputSection *, 0> sections; |
1383 | for (SectionCommand *cmd : sec->commands) { |
1384 | auto *isd = dyn_cast<InputSectionDescription>(Val: cmd); |
1385 | if (!isd) |
1386 | continue; |
1387 | bool hasLinkOrder = false; |
1388 | scriptSections.clear(); |
1389 | sections.clear(); |
1390 | for (InputSection *&isec : isd->sections) { |
1391 | if (isec->flags & SHF_LINK_ORDER) { |
1392 | InputSection *link = isec->getLinkOrderDep(); |
1393 | if (link && !link->getParent()) |
1394 | error(msg: toString(isec) + ": sh_link points to discarded section " + |
1395 | toString(link)); |
1396 | hasLinkOrder = true; |
1397 | } |
1398 | scriptSections.push_back(Elt: &isec); |
1399 | sections.push_back(Elt: isec); |
1400 | } |
1401 | if (hasLinkOrder && errorCount() == 0) { |
1402 | llvm::stable_sort(Range&: sections, C: compareByFilePosition); |
1403 | for (int i = 0, n = sections.size(); i != n; ++i) |
1404 | *scriptSections[i] = sections[i]; |
1405 | } |
1406 | } |
1407 | } |
1408 | } |
1409 | |
1410 | static void finalizeSynthetic(SyntheticSection *sec) { |
1411 | if (sec && sec->isNeeded() && sec->getParent()) { |
1412 | llvm::TimeTraceScope timeScope("Finalize synthetic sections" , sec->name); |
1413 | sec->finalizeContents(); |
1414 | } |
1415 | } |
1416 | |
1417 | // We need to generate and finalize the content that depends on the address of |
1418 | // InputSections. As the generation of the content may also alter InputSection |
1419 | // addresses we must converge to a fixed point. We do that here. See the comment |
1420 | // in Writer<ELFT>::finalizeSections(). |
1421 | template <class ELFT> void Writer<ELFT>::finalizeAddressDependentContent() { |
1422 | llvm::TimeTraceScope timeScope("Finalize address dependent content" ); |
1423 | ThunkCreator tc; |
1424 | AArch64Err843419Patcher a64p; |
1425 | ARMErr657417Patcher a32p; |
1426 | script->assignAddresses(); |
1427 | |
1428 | // .ARM.exidx and SHF_LINK_ORDER do not require precise addresses, but they |
1429 | // do require the relative addresses of OutputSections because linker scripts |
1430 | // can assign Virtual Addresses to OutputSections that are not monotonically |
1431 | // increasing. Anything here must be repeatable, since spilling may change |
1432 | // section order. |
1433 | const auto finalizeOrderDependentContent = [this] { |
1434 | for (Partition &part : partitions) |
1435 | finalizeSynthetic(sec: part.armExidx.get()); |
1436 | resolveShfLinkOrder(); |
1437 | }; |
1438 | finalizeOrderDependentContent(); |
1439 | |
1440 | // Converts call x@GDPLT to call __tls_get_addr |
1441 | if (config->emachine == EM_HEXAGON) |
1442 | hexagonTLSSymbolUpdate(outputSections); |
1443 | |
1444 | uint32_t pass = 0, assignPasses = 0; |
1445 | for (;;) { |
1446 | bool changed = target->needsThunks ? tc.createThunks(pass, outputSections) |
1447 | : target->relaxOnce(pass); |
1448 | bool spilled = script->spillSections(); |
1449 | changed |= spilled; |
1450 | ++pass; |
1451 | |
1452 | // With Thunk Size much smaller than branch range we expect to |
1453 | // converge quickly; if we get to 30 something has gone wrong. |
1454 | if (changed && pass >= 30) { |
1455 | error(msg: target->needsThunks ? "thunk creation not converged" |
1456 | : "relaxation not converged" ); |
1457 | break; |
1458 | } |
1459 | |
1460 | if (config->fixCortexA53Errata843419) { |
1461 | if (changed) |
1462 | script->assignAddresses(); |
1463 | changed |= a64p.createFixes(); |
1464 | } |
1465 | if (config->fixCortexA8) { |
1466 | if (changed) |
1467 | script->assignAddresses(); |
1468 | changed |= a32p.createFixes(); |
1469 | } |
1470 | |
1471 | finalizeSynthetic(sec: in.got.get()); |
1472 | if (in.mipsGot) |
1473 | in.mipsGot->updateAllocSize(); |
1474 | |
1475 | for (Partition &part : partitions) { |
1476 | // The R_AARCH64_AUTH_RELATIVE has a smaller addend field as bits [63:32] |
1477 | // encode the signing schema. We've put relocations in .relr.auth.dyn |
1478 | // during RelocationScanner::processAux, but the target VA for some of |
1479 | // them might be wider than 32 bits. We can only know the final VA at this |
1480 | // point, so move relocations with large values from .relr.auth.dyn to |
1481 | // .rela.dyn. See also AArch64::relocate. |
1482 | if (part.relrAuthDyn) { |
1483 | auto it = llvm::remove_if( |
1484 | part.relrAuthDyn->relocs, [&part](const RelativeReloc &elem) { |
1485 | const Relocation &reloc = elem.inputSec->relocs()[elem.relocIdx]; |
1486 | if (isInt<32>(x: reloc.sym->getVA(addend: reloc.addend))) |
1487 | return false; |
1488 | part.relaDyn->addReloc(reloc: {R_AARCH64_AUTH_RELATIVE, elem.inputSec, |
1489 | reloc.offset, |
1490 | DynamicReloc::AddendOnlyWithTargetVA, |
1491 | *reloc.sym, reloc.addend, R_ABS}); |
1492 | return true; |
1493 | }); |
1494 | changed |= (it != part.relrAuthDyn->relocs.end()); |
1495 | part.relrAuthDyn->relocs.erase(it, part.relrAuthDyn->relocs.end()); |
1496 | } |
1497 | if (part.relaDyn) |
1498 | changed |= part.relaDyn->updateAllocSize(); |
1499 | if (part.relrDyn) |
1500 | changed |= part.relrDyn->updateAllocSize(); |
1501 | if (part.relrAuthDyn) |
1502 | changed |= part.relrAuthDyn->updateAllocSize(); |
1503 | if (part.memtagGlobalDescriptors) |
1504 | changed |= part.memtagGlobalDescriptors->updateAllocSize(); |
1505 | } |
1506 | |
1507 | std::pair<const OutputSection *, const Defined *> changes = |
1508 | script->assignAddresses(); |
1509 | if (!changed) { |
1510 | // Some symbols may be dependent on section addresses. When we break the |
1511 | // loop, the symbol values are finalized because a previous |
1512 | // assignAddresses() finalized section addresses. |
1513 | if (!changes.first && !changes.second) |
1514 | break; |
1515 | if (++assignPasses == 5) { |
1516 | if (changes.first) |
1517 | errorOrWarn(msg: "address (0x" + Twine::utohexstr(Val: changes.first->addr) + |
1518 | ") of section '" + changes.first->name + |
1519 | "' does not converge" ); |
1520 | if (changes.second) |
1521 | errorOrWarn(msg: "assignment to symbol " + toString(*changes.second) + |
1522 | " does not converge" ); |
1523 | break; |
1524 | } |
1525 | } else if (spilled) { |
1526 | // Spilling can change relative section order. |
1527 | finalizeOrderDependentContent(); |
1528 | } |
1529 | } |
1530 | if (!config->relocatable) |
1531 | target->finalizeRelax(passes: pass); |
1532 | |
1533 | if (config->relocatable) |
1534 | for (OutputSection *sec : outputSections) |
1535 | sec->addr = 0; |
1536 | |
1537 | // If addrExpr is set, the address may not be a multiple of the alignment. |
1538 | // Warn because this is error-prone. |
1539 | for (SectionCommand *cmd : script->sectionCommands) |
1540 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) { |
1541 | OutputSection *osec = &osd->osec; |
1542 | if (osec->addr % osec->addralign != 0) |
1543 | warn(msg: "address (0x" + Twine::utohexstr(Val: osec->addr) + ") of section " + |
1544 | osec->name + " is not a multiple of alignment (" + |
1545 | Twine(osec->addralign) + ")" ); |
1546 | } |
1547 | |
1548 | // Sizes are no longer allowed to grow, so all allowable spills have been |
1549 | // taken. Remove any leftover potential spills. |
1550 | script->erasePotentialSpillSections(); |
1551 | } |
1552 | |
1553 | // If Input Sections have been shrunk (basic block sections) then |
1554 | // update symbol values and sizes associated with these sections. With basic |
1555 | // block sections, input sections can shrink when the jump instructions at |
1556 | // the end of the section are relaxed. |
1557 | static void fixSymbolsAfterShrinking() { |
1558 | for (InputFile *File : ctx.objectFiles) { |
1559 | parallelForEach(R: File->getSymbols(), Fn: [&](Symbol *Sym) { |
1560 | auto *def = dyn_cast<Defined>(Val: Sym); |
1561 | if (!def) |
1562 | return; |
1563 | |
1564 | const SectionBase *sec = def->section; |
1565 | if (!sec) |
1566 | return; |
1567 | |
1568 | const InputSectionBase *inputSec = dyn_cast<InputSectionBase>(Val: sec); |
1569 | if (!inputSec || !inputSec->bytesDropped) |
1570 | return; |
1571 | |
1572 | const size_t OldSize = inputSec->content().size(); |
1573 | const size_t NewSize = OldSize - inputSec->bytesDropped; |
1574 | |
1575 | if (def->value > NewSize && def->value <= OldSize) { |
1576 | LLVM_DEBUG(llvm::dbgs() |
1577 | << "Moving symbol " << Sym->getName() << " from " |
1578 | << def->value << " to " |
1579 | << def->value - inputSec->bytesDropped << " bytes\n" ); |
1580 | def->value -= inputSec->bytesDropped; |
1581 | return; |
1582 | } |
1583 | |
1584 | if (def->value + def->size > NewSize && def->value <= OldSize && |
1585 | def->value + def->size <= OldSize) { |
1586 | LLVM_DEBUG(llvm::dbgs() |
1587 | << "Shrinking symbol " << Sym->getName() << " from " |
1588 | << def->size << " to " << def->size - inputSec->bytesDropped |
1589 | << " bytes\n" ); |
1590 | def->size -= inputSec->bytesDropped; |
1591 | } |
1592 | }); |
1593 | } |
1594 | } |
1595 | |
1596 | // If basic block sections exist, there are opportunities to delete fall thru |
1597 | // jumps and shrink jump instructions after basic block reordering. This |
1598 | // relaxation pass does that. It is only enabled when --optimize-bb-jumps |
1599 | // option is used. |
1600 | template <class ELFT> void Writer<ELFT>::optimizeBasicBlockJumps() { |
1601 | assert(config->optimizeBBJumps); |
1602 | SmallVector<InputSection *, 0> storage; |
1603 | |
1604 | script->assignAddresses(); |
1605 | // For every output section that has executable input sections, this |
1606 | // does the following: |
1607 | // 1. Deletes all direct jump instructions in input sections that |
1608 | // jump to the following section as it is not required. |
1609 | // 2. If there are two consecutive jump instructions, it checks |
1610 | // if they can be flipped and one can be deleted. |
1611 | for (OutputSection *osec : outputSections) { |
1612 | if (!(osec->flags & SHF_EXECINSTR)) |
1613 | continue; |
1614 | ArrayRef<InputSection *> sections = getInputSections(os: *osec, storage); |
1615 | size_t numDeleted = 0; |
1616 | // Delete all fall through jump instructions. Also, check if two |
1617 | // consecutive jump instructions can be flipped so that a fall |
1618 | // through jmp instruction can be deleted. |
1619 | for (size_t i = 0, e = sections.size(); i != e; ++i) { |
1620 | InputSection *next = i + 1 < sections.size() ? sections[i + 1] : nullptr; |
1621 | InputSection &sec = *sections[i]; |
1622 | numDeleted += target->deleteFallThruJmpInsn(is&: sec, file: sec.file, nextIS: next); |
1623 | } |
1624 | if (numDeleted > 0) { |
1625 | script->assignAddresses(); |
1626 | LLVM_DEBUG(llvm::dbgs() |
1627 | << "Removing " << numDeleted << " fall through jumps\n" ); |
1628 | } |
1629 | } |
1630 | |
1631 | fixSymbolsAfterShrinking(); |
1632 | |
1633 | for (OutputSection *osec : outputSections) |
1634 | for (InputSection *is : getInputSections(os: *osec, storage)) |
1635 | is->trim(); |
1636 | } |
1637 | |
1638 | // In order to allow users to manipulate linker-synthesized sections, |
1639 | // we had to add synthetic sections to the input section list early, |
1640 | // even before we make decisions whether they are needed. This allows |
1641 | // users to write scripts like this: ".mygot : { .got }". |
1642 | // |
1643 | // Doing it has an unintended side effects. If it turns out that we |
1644 | // don't need a .got (for example) at all because there's no |
1645 | // relocation that needs a .got, we don't want to emit .got. |
1646 | // |
1647 | // To deal with the above problem, this function is called after |
1648 | // scanRelocations is called to remove synthetic sections that turn |
1649 | // out to be empty. |
1650 | static void removeUnusedSyntheticSections() { |
1651 | // All input synthetic sections that can be empty are placed after |
1652 | // all regular ones. Reverse iterate to find the first synthetic section |
1653 | // after a non-synthetic one which will be our starting point. |
1654 | auto start = |
1655 | llvm::find_if(Range: llvm::reverse(C&: ctx.inputSections), P: [](InputSectionBase *s) { |
1656 | return !isa<SyntheticSection>(Val: s); |
1657 | }).base(); |
1658 | |
1659 | // Remove unused synthetic sections from ctx.inputSections; |
1660 | DenseSet<InputSectionBase *> unused; |
1661 | auto end = |
1662 | std::remove_if(first: start, last: ctx.inputSections.end(), pred: [&](InputSectionBase *s) { |
1663 | auto *sec = cast<SyntheticSection>(Val: s); |
1664 | if (sec->getParent() && sec->isNeeded()) |
1665 | return false; |
1666 | // .relr.auth.dyn relocations may be moved to .rela.dyn in |
1667 | // finalizeAddressDependentContent, making .rela.dyn no longer empty. |
1668 | // Conservatively keep .rela.dyn. .relr.auth.dyn can be made empty, but |
1669 | // we would fail to remove it here. |
1670 | if (config->emachine == EM_AARCH64 && config->relrPackDynRelocs) |
1671 | if (auto *relSec = dyn_cast<RelocationBaseSection>(Val: sec)) |
1672 | if (relSec == mainPart->relaDyn.get()) |
1673 | return false; |
1674 | unused.insert(V: sec); |
1675 | return true; |
1676 | }); |
1677 | ctx.inputSections.erase(CS: end, CE: ctx.inputSections.end()); |
1678 | |
1679 | // Remove unused synthetic sections from the corresponding input section |
1680 | // description and orphanSections. |
1681 | for (auto *sec : unused) |
1682 | if (OutputSection *osec = cast<SyntheticSection>(Val: sec)->getParent()) |
1683 | for (SectionCommand *cmd : osec->commands) |
1684 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: cmd)) |
1685 | llvm::erase_if(C&: isd->sections, P: [&](InputSection *isec) { |
1686 | return unused.count(V: isec); |
1687 | }); |
1688 | llvm::erase_if(C&: script->orphanSections, P: [&](const InputSectionBase *sec) { |
1689 | return unused.count(V: sec); |
1690 | }); |
1691 | } |
1692 | |
1693 | // Create output section objects and add them to OutputSections. |
1694 | template <class ELFT> void Writer<ELFT>::finalizeSections() { |
1695 | if (!config->relocatable) { |
1696 | Out::preinitArray = findSection(name: ".preinit_array" ); |
1697 | Out::initArray = findSection(name: ".init_array" ); |
1698 | Out::finiArray = findSection(name: ".fini_array" ); |
1699 | |
1700 | // The linker needs to define SECNAME_start, SECNAME_end and SECNAME_stop |
1701 | // symbols for sections, so that the runtime can get the start and end |
1702 | // addresses of each section by section name. Add such symbols. |
1703 | addStartEndSymbols(); |
1704 | for (SectionCommand *cmd : script->sectionCommands) |
1705 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) |
1706 | addStartStopSymbols(osec&: osd->osec); |
1707 | |
1708 | // Add _DYNAMIC symbol. Unlike GNU gold, our _DYNAMIC symbol has no type. |
1709 | // It should be okay as no one seems to care about the type. |
1710 | // Even the author of gold doesn't remember why gold behaves that way. |
1711 | // https://sourceware.org/ml/binutils/2002-03/msg00360.html |
1712 | if (mainPart->dynamic->parent) { |
1713 | Symbol *s = symtab.addSymbol(newSym: Defined{ |
1714 | ctx.internalFile, "_DYNAMIC" , STB_WEAK, STV_HIDDEN, STT_NOTYPE, |
1715 | /*value=*/0, /*size=*/0, mainPart->dynamic.get()}); |
1716 | s->isUsedInRegularObj = true; |
1717 | } |
1718 | |
1719 | // Define __rel[a]_iplt_{start,end} symbols if needed. |
1720 | addRelIpltSymbols(); |
1721 | |
1722 | // RISC-V's gp can address +/- 2 KiB, set it to .sdata + 0x800. This symbol |
1723 | // should only be defined in an executable. If .sdata does not exist, its |
1724 | // value/section does not matter but it has to be relative, so set its |
1725 | // st_shndx arbitrarily to 1 (Out::elfHeader). |
1726 | if (config->emachine == EM_RISCV) { |
1727 | ElfSym::riscvGlobalPointer = nullptr; |
1728 | if (!config->shared) { |
1729 | OutputSection *sec = findSection(name: ".sdata" ); |
1730 | addOptionalRegular( |
1731 | name: "__global_pointer$" , sec: sec ? sec : Out::elfHeader, val: 0x800, stOther: STV_DEFAULT); |
1732 | // Set riscvGlobalPointer to be used by the optional global pointer |
1733 | // relaxation. |
1734 | if (config->relaxGP) { |
1735 | Symbol *s = symtab.find(name: "__global_pointer$" ); |
1736 | if (s && s->isDefined()) |
1737 | ElfSym::riscvGlobalPointer = cast<Defined>(Val: s); |
1738 | } |
1739 | } |
1740 | } |
1741 | |
1742 | if (config->emachine == EM_386 || config->emachine == EM_X86_64) { |
1743 | // On targets that support TLSDESC, _TLS_MODULE_BASE_ is defined in such a |
1744 | // way that: |
1745 | // |
1746 | // 1) Without relaxation: it produces a dynamic TLSDESC relocation that |
1747 | // computes 0. |
1748 | // 2) With LD->LE relaxation: _TLS_MODULE_BASE_@tpoff = 0 (lowest address |
1749 | // in the TLS block). |
1750 | // |
1751 | // 2) is special cased in @tpoff computation. To satisfy 1), we define it |
1752 | // as an absolute symbol of zero. This is different from GNU linkers which |
1753 | // define _TLS_MODULE_BASE_ relative to the first TLS section. |
1754 | Symbol *s = symtab.find(name: "_TLS_MODULE_BASE_" ); |
1755 | if (s && s->isUndefined()) { |
1756 | s->resolve(other: Defined{ctx.internalFile, StringRef(), STB_GLOBAL, |
1757 | STV_HIDDEN, STT_TLS, /*value=*/0, 0, |
1758 | /*section=*/nullptr}); |
1759 | ElfSym::tlsModuleBase = cast<Defined>(Val: s); |
1760 | } |
1761 | } |
1762 | |
1763 | // This responsible for splitting up .eh_frame section into |
1764 | // pieces. The relocation scan uses those pieces, so this has to be |
1765 | // earlier. |
1766 | { |
1767 | llvm::TimeTraceScope timeScope("Finalize .eh_frame" ); |
1768 | for (Partition &part : partitions) |
1769 | finalizeSynthetic(sec: part.ehFrame.get()); |
1770 | } |
1771 | } |
1772 | |
1773 | demoteSymbolsAndComputeIsPreemptible(); |
1774 | |
1775 | if (config->copyRelocs && config->discard != DiscardPolicy::None) |
1776 | markUsedLocalSymbols<ELFT>(); |
1777 | demoteAndCopyLocalSymbols(); |
1778 | |
1779 | if (config->copyRelocs) |
1780 | addSectionSymbols(); |
1781 | |
1782 | // Change values of linker-script-defined symbols from placeholders (assigned |
1783 | // by declareSymbols) to actual definitions. |
1784 | script->processSymbolAssignments(); |
1785 | |
1786 | if (!config->relocatable) { |
1787 | llvm::TimeTraceScope timeScope("Scan relocations" ); |
1788 | // Scan relocations. This must be done after every symbol is declared so |
1789 | // that we can correctly decide if a dynamic relocation is needed. This is |
1790 | // called after processSymbolAssignments() because it needs to know whether |
1791 | // a linker-script-defined symbol is absolute. |
1792 | ppc64noTocRelax.clear(); |
1793 | scanRelocations<ELFT>(); |
1794 | reportUndefinedSymbols(); |
1795 | postScanRelocations(); |
1796 | |
1797 | if (in.plt && in.plt->isNeeded()) |
1798 | in.plt->addSymbols(); |
1799 | if (in.iplt && in.iplt->isNeeded()) |
1800 | in.iplt->addSymbols(); |
1801 | |
1802 | if (config->unresolvedSymbolsInShlib != UnresolvedPolicy::Ignore) { |
1803 | auto diagnose = |
1804 | config->unresolvedSymbolsInShlib == UnresolvedPolicy::ReportError |
1805 | ? errorOrWarn |
1806 | : warn; |
1807 | // Error on undefined symbols in a shared object, if all of its DT_NEEDED |
1808 | // entries are seen. These cases would otherwise lead to runtime errors |
1809 | // reported by the dynamic linker. |
1810 | // |
1811 | // ld.bfd traces all DT_NEEDED to emulate the logic of the dynamic linker |
1812 | // to catch more cases. That is too much for us. Our approach resembles |
1813 | // the one used in ld.gold, achieves a good balance to be useful but not |
1814 | // too smart. |
1815 | // |
1816 | // If a DSO reference is resolved by a SharedSymbol, but the SharedSymbol |
1817 | // is overridden by a hidden visibility Defined (which is later discarded |
1818 | // due to GC), don't report the diagnostic. However, this may indicate an |
1819 | // unintended SharedSymbol. |
1820 | for (SharedFile *file : ctx.sharedFiles) { |
1821 | bool allNeededIsKnown = |
1822 | llvm::all_of(file->dtNeeded, [&](StringRef needed) { |
1823 | return symtab.soNames.count(Val: CachedHashStringRef(needed)); |
1824 | }); |
1825 | if (!allNeededIsKnown) |
1826 | continue; |
1827 | for (Symbol *sym : file->requiredSymbols) { |
1828 | if (sym->dsoDefined) |
1829 | continue; |
1830 | if (sym->isUndefined() && !sym->isWeak()) { |
1831 | diagnose("undefined reference: " + toString(*sym) + |
1832 | "\n>>> referenced by " + toString(f: file) + |
1833 | " (disallowed by --no-allow-shlib-undefined)" ); |
1834 | } else if (sym->isDefined() && sym->computeBinding() == STB_LOCAL) { |
1835 | diagnose("non-exported symbol '" + toString(*sym) + "' in '" + |
1836 | toString(f: sym->file) + "' is referenced by DSO '" + |
1837 | toString(f: file) + "'" ); |
1838 | } |
1839 | } |
1840 | } |
1841 | } |
1842 | } |
1843 | |
1844 | { |
1845 | llvm::TimeTraceScope timeScope("Add symbols to symtabs" ); |
1846 | // Now that we have defined all possible global symbols including linker- |
1847 | // synthesized ones. Visit all symbols to give the finishing touches. |
1848 | for (Symbol *sym : symtab.getSymbols()) { |
1849 | if (!sym->isUsedInRegularObj || !includeInSymtab(b: *sym)) |
1850 | continue; |
1851 | if (!config->relocatable) |
1852 | sym->binding = sym->computeBinding(); |
1853 | if (in.symTab) |
1854 | in.symTab->addSymbol(sym); |
1855 | |
1856 | if (sym->includeInDynsym()) { |
1857 | partitions[sym->partition - 1].dynSymTab->addSymbol(sym); |
1858 | if (auto *file = dyn_cast_or_null<SharedFile>(Val: sym->file)) |
1859 | if (file->isNeeded && !sym->isUndefined()) |
1860 | addVerneed(ss: sym); |
1861 | } |
1862 | } |
1863 | |
1864 | // We also need to scan the dynamic relocation tables of the other |
1865 | // partitions and add any referenced symbols to the partition's dynsym. |
1866 | for (Partition &part : MutableArrayRef<Partition>(partitions).slice(N: 1)) { |
1867 | DenseSet<Symbol *> syms; |
1868 | for (const SymbolTableEntry &e : part.dynSymTab->getSymbols()) |
1869 | syms.insert(V: e.sym); |
1870 | for (DynamicReloc &reloc : part.relaDyn->relocs) |
1871 | if (reloc.sym && reloc.needsDynSymIndex() && |
1872 | syms.insert(V: reloc.sym).second) |
1873 | part.dynSymTab->addSymbol(sym: reloc.sym); |
1874 | } |
1875 | } |
1876 | |
1877 | if (in.mipsGot) |
1878 | in.mipsGot->build(); |
1879 | |
1880 | removeUnusedSyntheticSections(); |
1881 | script->diagnoseOrphanHandling(); |
1882 | script->diagnoseMissingSGSectionAddress(); |
1883 | |
1884 | sortSections(); |
1885 | |
1886 | // Create a list of OutputSections, assign sectionIndex, and populate |
1887 | // in.shStrTab. |
1888 | for (SectionCommand *cmd : script->sectionCommands) |
1889 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) { |
1890 | OutputSection *osec = &osd->osec; |
1891 | outputSections.push_back(Elt: osec); |
1892 | osec->sectionIndex = outputSections.size(); |
1893 | osec->shName = in.shStrTab->addString(s: osec->name); |
1894 | } |
1895 | |
1896 | // Prefer command line supplied address over other constraints. |
1897 | for (OutputSection *sec : outputSections) { |
1898 | auto i = config->sectionStartMap.find(Key: sec->name); |
1899 | if (i != config->sectionStartMap.end()) |
1900 | sec->addrExpr = [=] { return i->second; }; |
1901 | } |
1902 | |
1903 | // With the outputSections available check for GDPLT relocations |
1904 | // and add __tls_get_addr symbol if needed. |
1905 | if (config->emachine == EM_HEXAGON && hexagonNeedsTLSSymbol(outputSections)) { |
1906 | Symbol *sym = |
1907 | symtab.addSymbol(newSym: Undefined{ctx.internalFile, "__tls_get_addr" , |
1908 | STB_GLOBAL, STV_DEFAULT, STT_NOTYPE}); |
1909 | sym->isPreemptible = true; |
1910 | partitions[0].dynSymTab->addSymbol(sym); |
1911 | } |
1912 | |
1913 | // This is a bit of a hack. A value of 0 means undef, so we set it |
1914 | // to 1 to make __ehdr_start defined. The section number is not |
1915 | // particularly relevant. |
1916 | Out::elfHeader->sectionIndex = 1; |
1917 | Out::elfHeader->size = sizeof(typename ELFT::Ehdr); |
1918 | |
1919 | // Binary and relocatable output does not have PHDRS. |
1920 | // The headers have to be created before finalize as that can influence the |
1921 | // image base and the dynamic section on mips includes the image base. |
1922 | if (!config->relocatable && !config->oFormatBinary) { |
1923 | for (Partition &part : partitions) { |
1924 | part.phdrs = script->hasPhdrsCommands() ? script->createPhdrs() |
1925 | : createPhdrs(part); |
1926 | if (config->emachine == EM_ARM) { |
1927 | // PT_ARM_EXIDX is the ARM EHABI equivalent of PT_GNU_EH_FRAME |
1928 | addPhdrForSection(part, shType: SHT_ARM_EXIDX, pType: PT_ARM_EXIDX, pFlags: PF_R); |
1929 | } |
1930 | if (config->emachine == EM_MIPS) { |
1931 | // Add separate segments for MIPS-specific sections. |
1932 | addPhdrForSection(part, shType: SHT_MIPS_REGINFO, pType: PT_MIPS_REGINFO, pFlags: PF_R); |
1933 | addPhdrForSection(part, shType: SHT_MIPS_OPTIONS, pType: PT_MIPS_OPTIONS, pFlags: PF_R); |
1934 | addPhdrForSection(part, shType: SHT_MIPS_ABIFLAGS, pType: PT_MIPS_ABIFLAGS, pFlags: PF_R); |
1935 | } |
1936 | if (config->emachine == EM_RISCV) |
1937 | addPhdrForSection(part, shType: SHT_RISCV_ATTRIBUTES, pType: PT_RISCV_ATTRIBUTES, |
1938 | pFlags: PF_R); |
1939 | } |
1940 | Out::programHeaders->size = sizeof(Elf_Phdr) * mainPart->phdrs.size(); |
1941 | |
1942 | // Find the TLS segment. This happens before the section layout loop so that |
1943 | // Android relocation packing can look up TLS symbol addresses. We only need |
1944 | // to care about the main partition here because all TLS symbols were moved |
1945 | // to the main partition (see MarkLive.cpp). |
1946 | for (PhdrEntry *p : mainPart->phdrs) |
1947 | if (p->p_type == PT_TLS) |
1948 | Out::tlsPhdr = p; |
1949 | } |
1950 | |
1951 | // Some symbols are defined in term of program headers. Now that we |
1952 | // have the headers, we can find out which sections they point to. |
1953 | setReservedSymbolSections(); |
1954 | |
1955 | if (script->noCrossRefs.size()) { |
1956 | llvm::TimeTraceScope timeScope("Check NOCROSSREFS" ); |
1957 | checkNoCrossRefs<ELFT>(); |
1958 | } |
1959 | |
1960 | { |
1961 | llvm::TimeTraceScope timeScope("Finalize synthetic sections" ); |
1962 | |
1963 | finalizeSynthetic(sec: in.bss.get()); |
1964 | finalizeSynthetic(sec: in.bssRelRo.get()); |
1965 | finalizeSynthetic(sec: in.symTabShndx.get()); |
1966 | finalizeSynthetic(sec: in.shStrTab.get()); |
1967 | finalizeSynthetic(sec: in.strTab.get()); |
1968 | finalizeSynthetic(sec: in.got.get()); |
1969 | finalizeSynthetic(sec: in.mipsGot.get()); |
1970 | finalizeSynthetic(sec: in.igotPlt.get()); |
1971 | finalizeSynthetic(sec: in.gotPlt.get()); |
1972 | finalizeSynthetic(sec: in.relaPlt.get()); |
1973 | finalizeSynthetic(sec: in.plt.get()); |
1974 | finalizeSynthetic(sec: in.iplt.get()); |
1975 | finalizeSynthetic(sec: in.ppc32Got2.get()); |
1976 | finalizeSynthetic(sec: in.partIndex.get()); |
1977 | |
1978 | // Dynamic section must be the last one in this list and dynamic |
1979 | // symbol table section (dynSymTab) must be the first one. |
1980 | for (Partition &part : partitions) { |
1981 | if (part.relaDyn) { |
1982 | part.relaDyn->mergeRels(); |
1983 | // Compute DT_RELACOUNT to be used by part.dynamic. |
1984 | part.relaDyn->partitionRels(); |
1985 | finalizeSynthetic(sec: part.relaDyn.get()); |
1986 | } |
1987 | if (part.relrDyn) { |
1988 | part.relrDyn->mergeRels(); |
1989 | finalizeSynthetic(sec: part.relrDyn.get()); |
1990 | } |
1991 | if (part.relrAuthDyn) { |
1992 | part.relrAuthDyn->mergeRels(); |
1993 | finalizeSynthetic(sec: part.relrAuthDyn.get()); |
1994 | } |
1995 | |
1996 | finalizeSynthetic(sec: part.dynSymTab.get()); |
1997 | finalizeSynthetic(sec: part.gnuHashTab.get()); |
1998 | finalizeSynthetic(sec: part.hashTab.get()); |
1999 | finalizeSynthetic(sec: part.verDef.get()); |
2000 | finalizeSynthetic(sec: part.ehFrameHdr.get()); |
2001 | finalizeSynthetic(sec: part.verSym.get()); |
2002 | finalizeSynthetic(sec: part.verNeed.get()); |
2003 | finalizeSynthetic(sec: part.dynamic.get()); |
2004 | } |
2005 | } |
2006 | |
2007 | if (!script->hasSectionsCommand && !config->relocatable) |
2008 | fixSectionAlignments(); |
2009 | |
2010 | // This is used to: |
2011 | // 1) Create "thunks": |
2012 | // Jump instructions in many ISAs have small displacements, and therefore |
2013 | // they cannot jump to arbitrary addresses in memory. For example, RISC-V |
2014 | // JAL instruction can target only +-1 MiB from PC. It is a linker's |
2015 | // responsibility to create and insert small pieces of code between |
2016 | // sections to extend the ranges if jump targets are out of range. Such |
2017 | // code pieces are called "thunks". |
2018 | // |
2019 | // We add thunks at this stage. We couldn't do this before this point |
2020 | // because this is the earliest point where we know sizes of sections and |
2021 | // their layouts (that are needed to determine if jump targets are in |
2022 | // range). |
2023 | // |
2024 | // 2) Update the sections. We need to generate content that depends on the |
2025 | // address of InputSections. For example, MIPS GOT section content or |
2026 | // android packed relocations sections content. |
2027 | // |
2028 | // 3) Assign the final values for the linker script symbols. Linker scripts |
2029 | // sometimes using forward symbol declarations. We want to set the correct |
2030 | // values. They also might change after adding the thunks. |
2031 | finalizeAddressDependentContent(); |
2032 | |
2033 | // All information needed for OutputSection part of Map file is available. |
2034 | if (errorCount()) |
2035 | return; |
2036 | |
2037 | { |
2038 | llvm::TimeTraceScope timeScope("Finalize synthetic sections" ); |
2039 | // finalizeAddressDependentContent may have added local symbols to the |
2040 | // static symbol table. |
2041 | finalizeSynthetic(sec: in.symTab.get()); |
2042 | finalizeSynthetic(sec: in.debugNames.get()); |
2043 | finalizeSynthetic(sec: in.ppc64LongBranchTarget.get()); |
2044 | finalizeSynthetic(sec: in.armCmseSGSection.get()); |
2045 | } |
2046 | |
2047 | // Relaxation to delete inter-basic block jumps created by basic block |
2048 | // sections. Run after in.symTab is finalized as optimizeBasicBlockJumps |
2049 | // can relax jump instructions based on symbol offset. |
2050 | if (config->optimizeBBJumps) |
2051 | optimizeBasicBlockJumps(); |
2052 | |
2053 | // Fill other section headers. The dynamic table is finalized |
2054 | // at the end because some tags like RELSZ depend on result |
2055 | // of finalizing other sections. |
2056 | for (OutputSection *sec : outputSections) |
2057 | sec->finalize(); |
2058 | |
2059 | script->checkFinalScriptConditions(); |
2060 | |
2061 | if (config->emachine == EM_ARM && !config->isLE && config->armBe8) { |
2062 | addArmInputSectionMappingSymbols(); |
2063 | sortArmMappingSymbols(); |
2064 | } |
2065 | } |
2066 | |
2067 | // Ensure data sections are not mixed with executable sections when |
2068 | // --execute-only is used. --execute-only make pages executable but not |
2069 | // readable. |
2070 | template <class ELFT> void Writer<ELFT>::checkExecuteOnly() { |
2071 | if (!config->executeOnly) |
2072 | return; |
2073 | |
2074 | SmallVector<InputSection *, 0> storage; |
2075 | for (OutputSection *osec : outputSections) |
2076 | if (osec->flags & SHF_EXECINSTR) |
2077 | for (InputSection *isec : getInputSections(os: *osec, storage)) |
2078 | if (!(isec->flags & SHF_EXECINSTR)) |
2079 | error(msg: "cannot place " + toString(isec) + " into " + |
2080 | toString(s: osec->name) + |
2081 | ": --execute-only does not support intermingling data and code" ); |
2082 | } |
2083 | |
2084 | // The linker is expected to define SECNAME_start and SECNAME_end |
2085 | // symbols for a few sections. This function defines them. |
2086 | template <class ELFT> void Writer<ELFT>::addStartEndSymbols() { |
2087 | // If the associated output section does not exist, there is ambiguity as to |
2088 | // how we define _start and _end symbols for an init/fini section. Users |
2089 | // expect no "undefined symbol" linker errors and loaders expect equal |
2090 | // st_value but do not particularly care whether the symbols are defined or |
2091 | // not. We retain the output section so that the section indexes will be |
2092 | // correct. |
2093 | auto define = [=](StringRef start, StringRef end, OutputSection *os) { |
2094 | if (os) { |
2095 | Defined *startSym = addOptionalRegular(name: start, sec: os, val: 0); |
2096 | Defined *stopSym = addOptionalRegular(name: end, sec: os, val: -1); |
2097 | if (startSym || stopSym) |
2098 | os->usedInExpression = true; |
2099 | } else { |
2100 | addOptionalRegular(name: start, sec: Out::elfHeader, val: 0); |
2101 | addOptionalRegular(name: end, sec: Out::elfHeader, val: 0); |
2102 | } |
2103 | }; |
2104 | |
2105 | define("__preinit_array_start" , "__preinit_array_end" , Out::preinitArray); |
2106 | define("__init_array_start" , "__init_array_end" , Out::initArray); |
2107 | define("__fini_array_start" , "__fini_array_end" , Out::finiArray); |
2108 | |
2109 | // As a special case, don't unnecessarily retain .ARM.exidx, which would |
2110 | // create an empty PT_ARM_EXIDX. |
2111 | if (OutputSection *sec = findSection(name: ".ARM.exidx" )) |
2112 | define("__exidx_start" , "__exidx_end" , sec); |
2113 | } |
2114 | |
2115 | // If a section name is valid as a C identifier (which is rare because of |
2116 | // the leading '.'), linkers are expected to define __start_<secname> and |
2117 | // __stop_<secname> symbols. They are at beginning and end of the section, |
2118 | // respectively. This is not requested by the ELF standard, but GNU ld and |
2119 | // gold provide the feature, and used by many programs. |
2120 | template <class ELFT> |
2121 | void Writer<ELFT>::addStartStopSymbols(OutputSection &osec) { |
2122 | StringRef s = osec.name; |
2123 | if (!isValidCIdentifier(s)) |
2124 | return; |
2125 | Defined *startSym = addOptionalRegular(name: saver().save(S: "__start_" + s), sec: &osec, val: 0, |
2126 | stOther: config->zStartStopVisibility); |
2127 | Defined *stopSym = addOptionalRegular(name: saver().save(S: "__stop_" + s), sec: &osec, val: -1, |
2128 | stOther: config->zStartStopVisibility); |
2129 | if (startSym || stopSym) |
2130 | osec.usedInExpression = true; |
2131 | } |
2132 | |
2133 | static bool needsPtLoad(OutputSection *sec) { |
2134 | if (!(sec->flags & SHF_ALLOC)) |
2135 | return false; |
2136 | |
2137 | // Don't allocate VA space for TLS NOBITS sections. The PT_TLS PHDR is |
2138 | // responsible for allocating space for them, not the PT_LOAD that |
2139 | // contains the TLS initialization image. |
2140 | if ((sec->flags & SHF_TLS) && sec->type == SHT_NOBITS) |
2141 | return false; |
2142 | return true; |
2143 | } |
2144 | |
2145 | // Adjust phdr flags according to certain options. |
2146 | static uint64_t computeFlags(uint64_t flags) { |
2147 | if (config->omagic) |
2148 | return PF_R | PF_W | PF_X; |
2149 | if (config->executeOnly && (flags & PF_X)) |
2150 | return flags & ~PF_R; |
2151 | return flags; |
2152 | } |
2153 | |
2154 | // Decide which program headers to create and which sections to include in each |
2155 | // one. |
2156 | template <class ELFT> |
2157 | SmallVector<PhdrEntry *, 0> Writer<ELFT>::createPhdrs(Partition &part) { |
2158 | SmallVector<PhdrEntry *, 0> ret; |
2159 | auto addHdr = [&](unsigned type, unsigned flags) -> PhdrEntry * { |
2160 | ret.push_back(Elt: make<PhdrEntry>(args&: type, args&: flags)); |
2161 | return ret.back(); |
2162 | }; |
2163 | |
2164 | unsigned partNo = part.getNumber(); |
2165 | bool isMain = partNo == 1; |
2166 | |
2167 | // Add the first PT_LOAD segment for regular output sections. |
2168 | uint64_t flags = computeFlags(flags: PF_R); |
2169 | PhdrEntry *load = nullptr; |
2170 | |
2171 | // nmagic or omagic output does not have PT_PHDR, PT_INTERP, or the readonly |
2172 | // PT_LOAD. |
2173 | if (!config->nmagic && !config->omagic) { |
2174 | // The first phdr entry is PT_PHDR which describes the program header |
2175 | // itself. |
2176 | if (isMain) |
2177 | addHdr(PT_PHDR, PF_R)->add(Out::programHeaders); |
2178 | else |
2179 | addHdr(PT_PHDR, PF_R)->add(part.programHeaders->getParent()); |
2180 | |
2181 | // PT_INTERP must be the second entry if exists. |
2182 | if (OutputSection *cmd = findSection(name: ".interp" , partition: partNo)) |
2183 | addHdr(PT_INTERP, cmd->getPhdrFlags())->add(cmd); |
2184 | |
2185 | // Add the headers. We will remove them if they don't fit. |
2186 | // In the other partitions the headers are ordinary sections, so they don't |
2187 | // need to be added here. |
2188 | if (isMain) { |
2189 | load = addHdr(PT_LOAD, flags); |
2190 | load->add(sec: Out::elfHeader); |
2191 | load->add(sec: Out::programHeaders); |
2192 | } |
2193 | } |
2194 | |
2195 | // PT_GNU_RELRO includes all sections that should be marked as |
2196 | // read-only by dynamic linker after processing relocations. |
2197 | // Current dynamic loaders only support one PT_GNU_RELRO PHDR, give |
2198 | // an error message if more than one PT_GNU_RELRO PHDR is required. |
2199 | PhdrEntry *relRo = make<PhdrEntry>(args: PT_GNU_RELRO, args: PF_R); |
2200 | bool inRelroPhdr = false; |
2201 | OutputSection *relroEnd = nullptr; |
2202 | for (OutputSection *sec : outputSections) { |
2203 | if (sec->partition != partNo || !needsPtLoad(sec)) |
2204 | continue; |
2205 | if (isRelroSection(sec)) { |
2206 | inRelroPhdr = true; |
2207 | if (!relroEnd) |
2208 | relRo->add(sec); |
2209 | else |
2210 | error(msg: "section: " + sec->name + " is not contiguous with other relro" + |
2211 | " sections" ); |
2212 | } else if (inRelroPhdr) { |
2213 | inRelroPhdr = false; |
2214 | relroEnd = sec; |
2215 | } |
2216 | } |
2217 | relRo->p_align = 1; |
2218 | |
2219 | for (OutputSection *sec : outputSections) { |
2220 | if (!needsPtLoad(sec)) |
2221 | continue; |
2222 | |
2223 | // Normally, sections in partitions other than the current partition are |
2224 | // ignored. But partition number 255 is a special case: it contains the |
2225 | // partition end marker (.part.end). It needs to be added to the main |
2226 | // partition so that a segment is created for it in the main partition, |
2227 | // which will cause the dynamic loader to reserve space for the other |
2228 | // partitions. |
2229 | if (sec->partition != partNo) { |
2230 | if (isMain && sec->partition == 255) |
2231 | addHdr(PT_LOAD, computeFlags(flags: sec->getPhdrFlags()))->add(sec); |
2232 | continue; |
2233 | } |
2234 | |
2235 | // Segments are contiguous memory regions that has the same attributes |
2236 | // (e.g. executable or writable). There is one phdr for each segment. |
2237 | // Therefore, we need to create a new phdr when the next section has |
2238 | // incompatible flags or is loaded at a discontiguous address or memory |
2239 | // region using AT or AT> linker script command, respectively. |
2240 | // |
2241 | // As an exception, we don't create a separate load segment for the ELF |
2242 | // headers, even if the first "real" output has an AT or AT> attribute. |
2243 | // |
2244 | // In addition, NOBITS sections should only be placed at the end of a LOAD |
2245 | // segment (since it's represented as p_filesz < p_memsz). If we have a |
2246 | // not-NOBITS section after a NOBITS, we create a new LOAD for the latter |
2247 | // even if flags match, so as not to require actually writing the |
2248 | // supposed-to-be-NOBITS section to the output file. (However, we cannot do |
2249 | // so when hasSectionsCommand, since we cannot introduce the extra alignment |
2250 | // needed to create a new LOAD) |
2251 | uint64_t newFlags = computeFlags(flags: sec->getPhdrFlags()); |
2252 | // When --no-rosegment is specified, RO and RX sections are compatible. |
2253 | uint32_t incompatible = flags ^ newFlags; |
2254 | if (config->singleRoRx && !(newFlags & PF_W)) |
2255 | incompatible &= ~PF_X; |
2256 | if (incompatible) |
2257 | load = nullptr; |
2258 | |
2259 | bool sameLMARegion = |
2260 | load && !sec->lmaExpr && sec->lmaRegion == load->firstSec->lmaRegion; |
2261 | if (load && sec != relroEnd && |
2262 | sec->memRegion == load->firstSec->memRegion && |
2263 | (sameLMARegion || load->lastSec == Out::programHeaders) && |
2264 | (script->hasSectionsCommand || sec->type == SHT_NOBITS || |
2265 | load->lastSec->type != SHT_NOBITS)) { |
2266 | load->p_flags |= newFlags; |
2267 | } else { |
2268 | load = addHdr(PT_LOAD, newFlags); |
2269 | flags = newFlags; |
2270 | } |
2271 | |
2272 | load->add(sec); |
2273 | } |
2274 | |
2275 | // Add a TLS segment if any. |
2276 | PhdrEntry *tlsHdr = make<PhdrEntry>(args: PT_TLS, args: PF_R); |
2277 | for (OutputSection *sec : outputSections) |
2278 | if (sec->partition == partNo && sec->flags & SHF_TLS) |
2279 | tlsHdr->add(sec); |
2280 | if (tlsHdr->firstSec) |
2281 | ret.push_back(Elt: tlsHdr); |
2282 | |
2283 | // Add an entry for .dynamic. |
2284 | if (OutputSection *sec = part.dynamic->getParent()) |
2285 | addHdr(PT_DYNAMIC, sec->getPhdrFlags())->add(sec); |
2286 | |
2287 | if (relRo->firstSec) |
2288 | ret.push_back(Elt: relRo); |
2289 | |
2290 | // PT_GNU_EH_FRAME is a special section pointing on .eh_frame_hdr. |
2291 | if (part.ehFrame->isNeeded() && part.ehFrameHdr && |
2292 | part.ehFrame->getParent() && part.ehFrameHdr->getParent()) |
2293 | addHdr(PT_GNU_EH_FRAME, part.ehFrameHdr->getParent()->getPhdrFlags()) |
2294 | ->add(part.ehFrameHdr->getParent()); |
2295 | |
2296 | if (config->osabi == ELFOSABI_OPENBSD) { |
2297 | // PT_OPENBSD_MUTABLE makes the dynamic linker fill the segment with |
2298 | // zero data, like bss, but it can be treated differently. |
2299 | if (OutputSection *cmd = findSection(name: ".openbsd.mutable" , partition: partNo)) |
2300 | addHdr(PT_OPENBSD_MUTABLE, cmd->getPhdrFlags())->add(cmd); |
2301 | |
2302 | // PT_OPENBSD_RANDOMIZE makes the dynamic linker fill the segment |
2303 | // with random data. |
2304 | if (OutputSection *cmd = findSection(name: ".openbsd.randomdata" , partition: partNo)) |
2305 | addHdr(PT_OPENBSD_RANDOMIZE, cmd->getPhdrFlags())->add(cmd); |
2306 | |
2307 | // PT_OPENBSD_SYSCALLS makes the kernel and dynamic linker register |
2308 | // system call sites. |
2309 | if (OutputSection *cmd = findSection(name: ".openbsd.syscalls" , partition: partNo)) |
2310 | addHdr(PT_OPENBSD_SYSCALLS, cmd->getPhdrFlags())->add(cmd); |
2311 | } |
2312 | |
2313 | if (config->zGnustack != GnuStackKind::None) { |
2314 | // PT_GNU_STACK is a special section to tell the loader to make the |
2315 | // pages for the stack non-executable. If you really want an executable |
2316 | // stack, you can pass -z execstack, but that's not recommended for |
2317 | // security reasons. |
2318 | unsigned perm = PF_R | PF_W; |
2319 | if (config->zGnustack == GnuStackKind::Exec) |
2320 | perm |= PF_X; |
2321 | addHdr(PT_GNU_STACK, perm)->p_memsz = config->zStackSize; |
2322 | } |
2323 | |
2324 | // PT_OPENBSD_WXNEEDED is a OpenBSD-specific header to mark the executable |
2325 | // is expected to perform W^X violations, such as calling mprotect(2) or |
2326 | // mmap(2) with PROT_WRITE | PROT_EXEC, which is prohibited by default on |
2327 | // OpenBSD. |
2328 | if (config->zWxneeded) |
2329 | addHdr(PT_OPENBSD_WXNEEDED, PF_X); |
2330 | |
2331 | if (OutputSection *cmd = findSection(name: ".note.gnu.property" , partition: partNo)) |
2332 | addHdr(PT_GNU_PROPERTY, PF_R)->add(cmd); |
2333 | |
2334 | // Create one PT_NOTE per a group of contiguous SHT_NOTE sections with the |
2335 | // same alignment. |
2336 | PhdrEntry *note = nullptr; |
2337 | for (OutputSection *sec : outputSections) { |
2338 | if (sec->partition != partNo) |
2339 | continue; |
2340 | if (sec->type == SHT_NOTE && (sec->flags & SHF_ALLOC)) { |
2341 | if (!note || sec->lmaExpr || note->lastSec->addralign != sec->addralign) |
2342 | note = addHdr(PT_NOTE, PF_R); |
2343 | note->add(sec); |
2344 | } else { |
2345 | note = nullptr; |
2346 | } |
2347 | } |
2348 | return ret; |
2349 | } |
2350 | |
2351 | template <class ELFT> |
2352 | void Writer<ELFT>::addPhdrForSection(Partition &part, unsigned shType, |
2353 | unsigned pType, unsigned pFlags) { |
2354 | unsigned partNo = part.getNumber(); |
2355 | auto i = llvm::find_if(outputSections, [=](OutputSection *cmd) { |
2356 | return cmd->partition == partNo && cmd->type == shType; |
2357 | }); |
2358 | if (i == outputSections.end()) |
2359 | return; |
2360 | |
2361 | PhdrEntry *entry = make<PhdrEntry>(args&: pType, args&: pFlags); |
2362 | entry->add(sec: *i); |
2363 | part.phdrs.push_back(Elt: entry); |
2364 | } |
2365 | |
2366 | // Place the first section of each PT_LOAD to a different page (of maxPageSize). |
2367 | // This is achieved by assigning an alignment expression to addrExpr of each |
2368 | // such section. |
2369 | template <class ELFT> void Writer<ELFT>::fixSectionAlignments() { |
2370 | const PhdrEntry *prev; |
2371 | auto pageAlign = [&](const PhdrEntry *p) { |
2372 | OutputSection *cmd = p->firstSec; |
2373 | if (!cmd) |
2374 | return; |
2375 | cmd->alignExpr = [align = cmd->addralign]() { return align; }; |
2376 | if (!cmd->addrExpr) { |
2377 | // Prefer advancing to align(dot, maxPageSize) + dot%maxPageSize to avoid |
2378 | // padding in the file contents. |
2379 | // |
2380 | // When -z separate-code is used we must not have any overlap in pages |
2381 | // between an executable segment and a non-executable segment. We align to |
2382 | // the next maximum page size boundary on transitions between executable |
2383 | // and non-executable segments. |
2384 | // |
2385 | // SHT_LLVM_PART_EHDR marks the start of a partition. The partition |
2386 | // sections will be extracted to a separate file. Align to the next |
2387 | // maximum page size boundary so that we can find the ELF header at the |
2388 | // start. We cannot benefit from overlapping p_offset ranges with the |
2389 | // previous segment anyway. |
2390 | if (config->zSeparate == SeparateSegmentKind::Loadable || |
2391 | (config->zSeparate == SeparateSegmentKind::Code && prev && |
2392 | (prev->p_flags & PF_X) != (p->p_flags & PF_X)) || |
2393 | cmd->type == SHT_LLVM_PART_EHDR) |
2394 | cmd->addrExpr = [] { |
2395 | return alignToPowerOf2(Value: script->getDot(), Align: config->maxPageSize); |
2396 | }; |
2397 | // PT_TLS is at the start of the first RW PT_LOAD. If `p` includes PT_TLS, |
2398 | // it must be the RW. Align to p_align(PT_TLS) to make sure |
2399 | // p_vaddr(PT_LOAD)%p_align(PT_LOAD) = 0. Otherwise, if |
2400 | // sh_addralign(.tdata) < sh_addralign(.tbss), we will set p_align(PT_TLS) |
2401 | // to sh_addralign(.tbss), while p_vaddr(PT_TLS)=p_vaddr(PT_LOAD) may not |
2402 | // be congruent to 0 modulo p_align(PT_TLS). |
2403 | // |
2404 | // Technically this is not required, but as of 2019, some dynamic loaders |
2405 | // don't handle p_vaddr%p_align != 0 correctly, e.g. glibc (i386 and |
2406 | // x86-64) doesn't make runtime address congruent to p_vaddr modulo |
2407 | // p_align for dynamic TLS blocks (PR/24606), FreeBSD rtld has the same |
2408 | // bug, musl (TLS Variant 1 architectures) before 1.1.23 handled TLS |
2409 | // blocks correctly. We need to keep the workaround for a while. |
2410 | else if (Out::tlsPhdr && Out::tlsPhdr->firstSec == p->firstSec) |
2411 | cmd->addrExpr = [] { |
2412 | return alignToPowerOf2(Value: script->getDot(), Align: config->maxPageSize) + |
2413 | alignToPowerOf2(Value: script->getDot() % config->maxPageSize, |
2414 | Align: Out::tlsPhdr->p_align); |
2415 | }; |
2416 | else |
2417 | cmd->addrExpr = [] { |
2418 | return alignToPowerOf2(Value: script->getDot(), Align: config->maxPageSize) + |
2419 | script->getDot() % config->maxPageSize; |
2420 | }; |
2421 | } |
2422 | }; |
2423 | |
2424 | for (Partition &part : partitions) { |
2425 | prev = nullptr; |
2426 | for (const PhdrEntry *p : part.phdrs) |
2427 | if (p->p_type == PT_LOAD && p->firstSec) { |
2428 | pageAlign(p); |
2429 | prev = p; |
2430 | } |
2431 | } |
2432 | } |
2433 | |
2434 | // Compute an in-file position for a given section. The file offset must be the |
2435 | // same with its virtual address modulo the page size, so that the loader can |
2436 | // load executables without any address adjustment. |
2437 | static uint64_t computeFileOffset(OutputSection *os, uint64_t off) { |
2438 | // The first section in a PT_LOAD has to have congruent offset and address |
2439 | // modulo the maximum page size. |
2440 | if (os->ptLoad && os->ptLoad->firstSec == os) |
2441 | return alignTo(Value: off, Align: os->ptLoad->p_align, Skew: os->addr); |
2442 | |
2443 | // File offsets are not significant for .bss sections other than the first one |
2444 | // in a PT_LOAD/PT_TLS. By convention, we keep section offsets monotonically |
2445 | // increasing rather than setting to zero. |
2446 | if (os->type == SHT_NOBITS && |
2447 | (!Out::tlsPhdr || Out::tlsPhdr->firstSec != os)) |
2448 | return off; |
2449 | |
2450 | // If the section is not in a PT_LOAD, we just have to align it. |
2451 | if (!os->ptLoad) |
2452 | return alignToPowerOf2(Value: off, Align: os->addralign); |
2453 | |
2454 | // If two sections share the same PT_LOAD the file offset is calculated |
2455 | // using this formula: Off2 = Off1 + (VA2 - VA1). |
2456 | OutputSection *first = os->ptLoad->firstSec; |
2457 | return first->offset + os->addr - first->addr; |
2458 | } |
2459 | |
2460 | template <class ELFT> void Writer<ELFT>::assignFileOffsetsBinary() { |
2461 | // Compute the minimum LMA of all non-empty non-NOBITS sections as minAddr. |
2462 | auto needsOffset = [](OutputSection &sec) { |
2463 | return sec.type != SHT_NOBITS && (sec.flags & SHF_ALLOC) && sec.size > 0; |
2464 | }; |
2465 | uint64_t minAddr = UINT64_MAX; |
2466 | for (OutputSection *sec : outputSections) |
2467 | if (needsOffset(*sec)) { |
2468 | sec->offset = sec->getLMA(); |
2469 | minAddr = std::min(a: minAddr, b: sec->offset); |
2470 | } |
2471 | |
2472 | // Sections are laid out at LMA minus minAddr. |
2473 | fileSize = 0; |
2474 | for (OutputSection *sec : outputSections) |
2475 | if (needsOffset(*sec)) { |
2476 | sec->offset -= minAddr; |
2477 | fileSize = std::max(a: fileSize, b: sec->offset + sec->size); |
2478 | } |
2479 | } |
2480 | |
2481 | static std::string rangeToString(uint64_t addr, uint64_t len) { |
2482 | return "[0x" + utohexstr(X: addr) + ", 0x" + utohexstr(X: addr + len - 1) + "]" ; |
2483 | } |
2484 | |
2485 | // Assign file offsets to output sections. |
2486 | template <class ELFT> void Writer<ELFT>::assignFileOffsets() { |
2487 | Out::programHeaders->offset = Out::elfHeader->size; |
2488 | uint64_t off = Out::elfHeader->size + Out::programHeaders->size; |
2489 | |
2490 | PhdrEntry *lastRX = nullptr; |
2491 | for (Partition &part : partitions) |
2492 | for (PhdrEntry *p : part.phdrs) |
2493 | if (p->p_type == PT_LOAD && (p->p_flags & PF_X)) |
2494 | lastRX = p; |
2495 | |
2496 | // Layout SHF_ALLOC sections before non-SHF_ALLOC sections. A non-SHF_ALLOC |
2497 | // will not occupy file offsets contained by a PT_LOAD. |
2498 | for (OutputSection *sec : outputSections) { |
2499 | if (!(sec->flags & SHF_ALLOC)) |
2500 | continue; |
2501 | off = computeFileOffset(os: sec, off); |
2502 | sec->offset = off; |
2503 | if (sec->type != SHT_NOBITS) |
2504 | off += sec->size; |
2505 | |
2506 | // If this is a last section of the last executable segment and that |
2507 | // segment is the last loadable segment, align the offset of the |
2508 | // following section to avoid loading non-segments parts of the file. |
2509 | if (config->zSeparate != SeparateSegmentKind::None && lastRX && |
2510 | lastRX->lastSec == sec) |
2511 | off = alignToPowerOf2(Value: off, Align: config->maxPageSize); |
2512 | } |
2513 | for (OutputSection *osec : outputSections) { |
2514 | if (osec->flags & SHF_ALLOC) |
2515 | continue; |
2516 | osec->offset = alignToPowerOf2(Value: off, Align: osec->addralign); |
2517 | off = osec->offset + osec->size; |
2518 | } |
2519 | |
2520 | sectionHeaderOff = alignToPowerOf2(Value: off, Align: config->wordsize); |
2521 | fileSize = sectionHeaderOff + (outputSections.size() + 1) * sizeof(Elf_Shdr); |
2522 | |
2523 | // Our logic assumes that sections have rising VA within the same segment. |
2524 | // With use of linker scripts it is possible to violate this rule and get file |
2525 | // offset overlaps or overflows. That should never happen with a valid script |
2526 | // which does not move the location counter backwards and usually scripts do |
2527 | // not do that. Unfortunately, there are apps in the wild, for example, Linux |
2528 | // kernel, which control segment distribution explicitly and move the counter |
2529 | // backwards, so we have to allow doing that to support linking them. We |
2530 | // perform non-critical checks for overlaps in checkSectionOverlap(), but here |
2531 | // we want to prevent file size overflows because it would crash the linker. |
2532 | for (OutputSection *sec : outputSections) { |
2533 | if (sec->type == SHT_NOBITS) |
2534 | continue; |
2535 | if ((sec->offset > fileSize) || (sec->offset + sec->size > fileSize)) |
2536 | error(msg: "unable to place section " + sec->name + " at file offset " + |
2537 | rangeToString(addr: sec->offset, len: sec->size) + |
2538 | "; check your linker script for overflows" ); |
2539 | } |
2540 | } |
2541 | |
2542 | // Finalize the program headers. We call this function after we assign |
2543 | // file offsets and VAs to all sections. |
2544 | template <class ELFT> void Writer<ELFT>::setPhdrs(Partition &part) { |
2545 | for (PhdrEntry *p : part.phdrs) { |
2546 | OutputSection *first = p->firstSec; |
2547 | OutputSection *last = p->lastSec; |
2548 | |
2549 | // .ARM.exidx sections may not be within a single .ARM.exidx |
2550 | // output section. We always want to describe just the |
2551 | // SyntheticSection. |
2552 | if (part.armExidx && p->p_type == PT_ARM_EXIDX) { |
2553 | p->p_filesz = part.armExidx->getSize(); |
2554 | p->p_memsz = part.armExidx->getSize(); |
2555 | p->p_offset = first->offset + part.armExidx->outSecOff; |
2556 | p->p_vaddr = first->addr + part.armExidx->outSecOff; |
2557 | p->p_align = part.armExidx->addralign; |
2558 | if (part.elfHeader) |
2559 | p->p_offset -= part.elfHeader->getParent()->offset; |
2560 | |
2561 | if (!p->hasLMA) |
2562 | p->p_paddr = first->getLMA() + part.armExidx->outSecOff; |
2563 | return; |
2564 | } |
2565 | |
2566 | if (first) { |
2567 | p->p_filesz = last->offset - first->offset; |
2568 | if (last->type != SHT_NOBITS) |
2569 | p->p_filesz += last->size; |
2570 | |
2571 | p->p_memsz = last->addr + last->size - first->addr; |
2572 | p->p_offset = first->offset; |
2573 | p->p_vaddr = first->addr; |
2574 | |
2575 | // File offsets in partitions other than the main partition are relative |
2576 | // to the offset of the ELF headers. Perform that adjustment now. |
2577 | if (part.elfHeader) |
2578 | p->p_offset -= part.elfHeader->getParent()->offset; |
2579 | |
2580 | if (!p->hasLMA) |
2581 | p->p_paddr = first->getLMA(); |
2582 | } |
2583 | } |
2584 | } |
2585 | |
2586 | // A helper struct for checkSectionOverlap. |
2587 | namespace { |
2588 | struct SectionOffset { |
2589 | OutputSection *sec; |
2590 | uint64_t offset; |
2591 | }; |
2592 | } // namespace |
2593 | |
2594 | // Check whether sections overlap for a specific address range (file offsets, |
2595 | // load and virtual addresses). |
2596 | static void checkOverlap(StringRef name, std::vector<SectionOffset> §ions, |
2597 | bool isVirtualAddr) { |
2598 | llvm::sort(C&: sections, Comp: [=](const SectionOffset &a, const SectionOffset &b) { |
2599 | return a.offset < b.offset; |
2600 | }); |
2601 | |
2602 | // Finding overlap is easy given a vector is sorted by start position. |
2603 | // If an element starts before the end of the previous element, they overlap. |
2604 | for (size_t i = 1, end = sections.size(); i < end; ++i) { |
2605 | SectionOffset a = sections[i - 1]; |
2606 | SectionOffset b = sections[i]; |
2607 | if (b.offset >= a.offset + a.sec->size) |
2608 | continue; |
2609 | |
2610 | // If both sections are in OVERLAY we allow the overlapping of virtual |
2611 | // addresses, because it is what OVERLAY was designed for. |
2612 | if (isVirtualAddr && a.sec->inOverlay && b.sec->inOverlay) |
2613 | continue; |
2614 | |
2615 | errorOrWarn(msg: "section " + a.sec->name + " " + name + |
2616 | " range overlaps with " + b.sec->name + "\n>>> " + a.sec->name + |
2617 | " range is " + rangeToString(addr: a.offset, len: a.sec->size) + "\n>>> " + |
2618 | b.sec->name + " range is " + |
2619 | rangeToString(addr: b.offset, len: b.sec->size)); |
2620 | } |
2621 | } |
2622 | |
2623 | // Check for overlapping sections and address overflows. |
2624 | // |
2625 | // In this function we check that none of the output sections have overlapping |
2626 | // file offsets. For SHF_ALLOC sections we also check that the load address |
2627 | // ranges and the virtual address ranges don't overlap |
2628 | template <class ELFT> void Writer<ELFT>::checkSections() { |
2629 | // First, check that section's VAs fit in available address space for target. |
2630 | for (OutputSection *os : outputSections) |
2631 | if ((os->addr + os->size < os->addr) || |
2632 | (!ELFT::Is64Bits && os->addr + os->size > uint64_t(UINT32_MAX) + 1)) |
2633 | errorOrWarn(msg: "section " + os->name + " at 0x" + utohexstr(X: os->addr) + |
2634 | " of size 0x" + utohexstr(X: os->size) + |
2635 | " exceeds available address space" ); |
2636 | |
2637 | // Check for overlapping file offsets. In this case we need to skip any |
2638 | // section marked as SHT_NOBITS. These sections don't actually occupy space in |
2639 | // the file so Sec->Offset + Sec->Size can overlap with others. If --oformat |
2640 | // binary is specified only add SHF_ALLOC sections are added to the output |
2641 | // file so we skip any non-allocated sections in that case. |
2642 | std::vector<SectionOffset> fileOffs; |
2643 | for (OutputSection *sec : outputSections) |
2644 | if (sec->size > 0 && sec->type != SHT_NOBITS && |
2645 | (!config->oFormatBinary || (sec->flags & SHF_ALLOC))) |
2646 | fileOffs.push_back(x: {.sec: sec, .offset: sec->offset}); |
2647 | checkOverlap(name: "file" , sections&: fileOffs, isVirtualAddr: false); |
2648 | |
2649 | // When linking with -r there is no need to check for overlapping virtual/load |
2650 | // addresses since those addresses will only be assigned when the final |
2651 | // executable/shared object is created. |
2652 | if (config->relocatable) |
2653 | return; |
2654 | |
2655 | // Checking for overlapping virtual and load addresses only needs to take |
2656 | // into account SHF_ALLOC sections since others will not be loaded. |
2657 | // Furthermore, we also need to skip SHF_TLS sections since these will be |
2658 | // mapped to other addresses at runtime and can therefore have overlapping |
2659 | // ranges in the file. |
2660 | std::vector<SectionOffset> vmas; |
2661 | for (OutputSection *sec : outputSections) |
2662 | if (sec->size > 0 && (sec->flags & SHF_ALLOC) && !(sec->flags & SHF_TLS)) |
2663 | vmas.push_back(x: {.sec: sec, .offset: sec->addr}); |
2664 | checkOverlap(name: "virtual address" , sections&: vmas, isVirtualAddr: true); |
2665 | |
2666 | // Finally, check that the load addresses don't overlap. This will usually be |
2667 | // the same as the virtual addresses but can be different when using a linker |
2668 | // script with AT(). |
2669 | std::vector<SectionOffset> lmas; |
2670 | for (OutputSection *sec : outputSections) |
2671 | if (sec->size > 0 && (sec->flags & SHF_ALLOC) && !(sec->flags & SHF_TLS)) |
2672 | lmas.push_back(x: {.sec: sec, .offset: sec->getLMA()}); |
2673 | checkOverlap(name: "load address" , sections&: lmas, isVirtualAddr: false); |
2674 | } |
2675 | |
2676 | // The entry point address is chosen in the following ways. |
2677 | // |
2678 | // 1. the '-e' entry command-line option; |
2679 | // 2. the ENTRY(symbol) command in a linker control script; |
2680 | // 3. the value of the symbol _start, if present; |
2681 | // 4. the number represented by the entry symbol, if it is a number; |
2682 | // 5. the address 0. |
2683 | static uint64_t getEntryAddr() { |
2684 | // Case 1, 2 or 3 |
2685 | if (Symbol *b = symtab.find(name: config->entry)) |
2686 | return b->getVA(); |
2687 | |
2688 | // Case 4 |
2689 | uint64_t addr; |
2690 | if (to_integer(S: config->entry, Num&: addr)) |
2691 | return addr; |
2692 | |
2693 | // Case 5 |
2694 | if (config->warnMissingEntry) |
2695 | warn(msg: "cannot find entry symbol " + config->entry + |
2696 | "; not setting start address" ); |
2697 | return 0; |
2698 | } |
2699 | |
2700 | static uint16_t getELFType() { |
2701 | if (config->isPic) |
2702 | return ET_DYN; |
2703 | if (config->relocatable) |
2704 | return ET_REL; |
2705 | return ET_EXEC; |
2706 | } |
2707 | |
2708 | template <class ELFT> void Writer<ELFT>::() { |
2709 | writeEhdr<ELFT>(Out::bufferStart, *mainPart); |
2710 | writePhdrs<ELFT>(Out::bufferStart + sizeof(Elf_Ehdr), *mainPart); |
2711 | |
2712 | auto *eHdr = reinterpret_cast<Elf_Ehdr *>(Out::bufferStart); |
2713 | eHdr->e_type = getELFType(); |
2714 | eHdr->e_entry = getEntryAddr(); |
2715 | eHdr->e_shoff = sectionHeaderOff; |
2716 | |
2717 | // Write the section header table. |
2718 | // |
2719 | // The ELF header can only store numbers up to SHN_LORESERVE in the e_shnum |
2720 | // and e_shstrndx fields. When the value of one of these fields exceeds |
2721 | // SHN_LORESERVE ELF requires us to put sentinel values in the ELF header and |
2722 | // use fields in the section header at index 0 to store |
2723 | // the value. The sentinel values and fields are: |
2724 | // e_shnum = 0, SHdrs[0].sh_size = number of sections. |
2725 | // e_shstrndx = SHN_XINDEX, SHdrs[0].sh_link = .shstrtab section index. |
2726 | auto *sHdrs = reinterpret_cast<Elf_Shdr *>(Out::bufferStart + eHdr->e_shoff); |
2727 | size_t num = outputSections.size() + 1; |
2728 | if (num >= SHN_LORESERVE) |
2729 | sHdrs->sh_size = num; |
2730 | else |
2731 | eHdr->e_shnum = num; |
2732 | |
2733 | uint32_t strTabIndex = in.shStrTab->getParent()->sectionIndex; |
2734 | if (strTabIndex >= SHN_LORESERVE) { |
2735 | sHdrs->sh_link = strTabIndex; |
2736 | eHdr->e_shstrndx = SHN_XINDEX; |
2737 | } else { |
2738 | eHdr->e_shstrndx = strTabIndex; |
2739 | } |
2740 | |
2741 | for (OutputSection *sec : outputSections) |
2742 | sec->writeHeaderTo<ELFT>(++sHdrs); |
2743 | } |
2744 | |
2745 | // Open a result file. |
2746 | template <class ELFT> void Writer<ELFT>::openFile() { |
2747 | uint64_t maxSize = config->is64 ? INT64_MAX : UINT32_MAX; |
2748 | if (fileSize != size_t(fileSize) || maxSize < fileSize) { |
2749 | std::string msg; |
2750 | raw_string_ostream s(msg); |
2751 | s << "output file too large: " << Twine(fileSize) << " bytes\n" |
2752 | << "section sizes:\n" ; |
2753 | for (OutputSection *os : outputSections) |
2754 | s << os->name << ' ' << os->size << "\n" ; |
2755 | error(msg: s.str()); |
2756 | return; |
2757 | } |
2758 | |
2759 | unlinkAsync(path: config->outputFile); |
2760 | unsigned flags = 0; |
2761 | if (!config->relocatable) |
2762 | flags |= FileOutputBuffer::F_executable; |
2763 | if (!config->mmapOutputFile) |
2764 | flags |= FileOutputBuffer::F_no_mmap; |
2765 | Expected<std::unique_ptr<FileOutputBuffer>> bufferOrErr = |
2766 | FileOutputBuffer::create(FilePath: config->outputFile, Size: fileSize, Flags: flags); |
2767 | |
2768 | if (!bufferOrErr) { |
2769 | error(msg: "failed to open " + config->outputFile + ": " + |
2770 | llvm::toString(E: bufferOrErr.takeError())); |
2771 | return; |
2772 | } |
2773 | buffer = std::move(*bufferOrErr); |
2774 | Out::bufferStart = buffer->getBufferStart(); |
2775 | } |
2776 | |
2777 | template <class ELFT> void Writer<ELFT>::writeSectionsBinary() { |
2778 | parallel::TaskGroup tg; |
2779 | for (OutputSection *sec : outputSections) |
2780 | if (sec->flags & SHF_ALLOC) |
2781 | sec->writeTo<ELFT>(Out::bufferStart + sec->offset, tg); |
2782 | } |
2783 | |
2784 | static void fillTrap(uint8_t *i, uint8_t *end) { |
2785 | for (; i + 4 <= end; i += 4) |
2786 | memcpy(dest: i, src: &target->trapInstr, n: 4); |
2787 | } |
2788 | |
2789 | // Fill the last page of executable segments with trap instructions |
2790 | // instead of leaving them as zero. Even though it is not required by any |
2791 | // standard, it is in general a good thing to do for security reasons. |
2792 | // |
2793 | // We'll leave other pages in segments as-is because the rest will be |
2794 | // overwritten by output sections. |
2795 | template <class ELFT> void Writer<ELFT>::writeTrapInstr() { |
2796 | for (Partition &part : partitions) { |
2797 | // Fill the last page. |
2798 | for (PhdrEntry *p : part.phdrs) |
2799 | if (p->p_type == PT_LOAD && (p->p_flags & PF_X)) |
2800 | fillTrap(i: Out::bufferStart + |
2801 | alignDown(Value: p->firstSec->offset + p->p_filesz, Align: 4), |
2802 | end: Out::bufferStart + |
2803 | alignToPowerOf2(Value: p->firstSec->offset + p->p_filesz, |
2804 | Align: config->maxPageSize)); |
2805 | |
2806 | // Round up the file size of the last segment to the page boundary iff it is |
2807 | // an executable segment to ensure that other tools don't accidentally |
2808 | // trim the instruction padding (e.g. when stripping the file). |
2809 | PhdrEntry *last = nullptr; |
2810 | for (PhdrEntry *p : part.phdrs) |
2811 | if (p->p_type == PT_LOAD) |
2812 | last = p; |
2813 | |
2814 | if (last && (last->p_flags & PF_X)) |
2815 | last->p_memsz = last->p_filesz = |
2816 | alignToPowerOf2(Value: last->p_filesz, Align: config->maxPageSize); |
2817 | } |
2818 | } |
2819 | |
2820 | // Write section contents to a mmap'ed file. |
2821 | template <class ELFT> void Writer<ELFT>::writeSections() { |
2822 | llvm::TimeTraceScope timeScope("Write sections" ); |
2823 | |
2824 | { |
2825 | // In -r or --emit-relocs mode, write the relocation sections first as in |
2826 | // ELf_Rel targets we might find out that we need to modify the relocated |
2827 | // section while doing it. |
2828 | parallel::TaskGroup tg; |
2829 | for (OutputSection *sec : outputSections) |
2830 | if (isStaticRelSecType(type: sec->type)) |
2831 | sec->writeTo<ELFT>(Out::bufferStart + sec->offset, tg); |
2832 | } |
2833 | { |
2834 | parallel::TaskGroup tg; |
2835 | for (OutputSection *sec : outputSections) |
2836 | if (!isStaticRelSecType(type: sec->type)) |
2837 | sec->writeTo<ELFT>(Out::bufferStart + sec->offset, tg); |
2838 | } |
2839 | |
2840 | // Finally, check that all dynamic relocation addends were written correctly. |
2841 | if (config->checkDynamicRelocs && config->writeAddends) { |
2842 | for (OutputSection *sec : outputSections) |
2843 | if (isStaticRelSecType(type: sec->type)) |
2844 | sec->checkDynRelAddends(bufStart: Out::bufferStart); |
2845 | } |
2846 | } |
2847 | |
2848 | // Computes a hash value of Data using a given hash function. |
2849 | // In order to utilize multiple cores, we first split data into 1MB |
2850 | // chunks, compute a hash for each chunk, and then compute a hash value |
2851 | // of the hash values. |
2852 | static void |
2853 | computeHash(llvm::MutableArrayRef<uint8_t> hashBuf, |
2854 | llvm::ArrayRef<uint8_t> data, |
2855 | std::function<void(uint8_t *dest, ArrayRef<uint8_t> arr)> hashFn) { |
2856 | std::vector<ArrayRef<uint8_t>> chunks = split(arr: data, chunkSize: 1024 * 1024); |
2857 | const size_t hashesSize = chunks.size() * hashBuf.size(); |
2858 | std::unique_ptr<uint8_t[]> hashes(new uint8_t[hashesSize]); |
2859 | |
2860 | // Compute hash values. |
2861 | parallelFor(Begin: 0, End: chunks.size(), Fn: [&](size_t i) { |
2862 | hashFn(hashes.get() + i * hashBuf.size(), chunks[i]); |
2863 | }); |
2864 | |
2865 | // Write to the final output buffer. |
2866 | hashFn(hashBuf.data(), ArrayRef(hashes.get(), hashesSize)); |
2867 | } |
2868 | |
2869 | template <class ELFT> void Writer<ELFT>::writeBuildId() { |
2870 | if (!mainPart->buildId || !mainPart->buildId->getParent()) |
2871 | return; |
2872 | |
2873 | if (config->buildId == BuildIdKind::Hexstring) { |
2874 | for (Partition &part : partitions) |
2875 | part.buildId->writeBuildId(buf: config->buildIdVector); |
2876 | return; |
2877 | } |
2878 | |
2879 | // Compute a hash of all sections of the output file. |
2880 | size_t hashSize = mainPart->buildId->hashSize; |
2881 | std::unique_ptr<uint8_t[]> buildId(new uint8_t[hashSize]); |
2882 | MutableArrayRef<uint8_t> output(buildId.get(), hashSize); |
2883 | llvm::ArrayRef<uint8_t> input{Out::bufferStart, size_t(fileSize)}; |
2884 | |
2885 | // Fedora introduced build ID as "approximation of true uniqueness across all |
2886 | // binaries that might be used by overlapping sets of people". It does not |
2887 | // need some security goals that some hash algorithms strive to provide, e.g. |
2888 | // (second-)preimage and collision resistance. In practice people use 'md5' |
2889 | // and 'sha1' just for different lengths. Implement them with the more |
2890 | // efficient BLAKE3. |
2891 | switch (config->buildId) { |
2892 | case BuildIdKind::Fast: |
2893 | computeHash(output, input, [](uint8_t *dest, ArrayRef<uint8_t> arr) { |
2894 | write64le(P: dest, V: xxh3_64bits(data: arr)); |
2895 | }); |
2896 | break; |
2897 | case BuildIdKind::Md5: |
2898 | computeHash(output, input, [&](uint8_t *dest, ArrayRef<uint8_t> arr) { |
2899 | memcpy(dest: dest, src: BLAKE3::hash<16>(Data: arr).data(), n: hashSize); |
2900 | }); |
2901 | break; |
2902 | case BuildIdKind::Sha1: |
2903 | computeHash(output, input, [&](uint8_t *dest, ArrayRef<uint8_t> arr) { |
2904 | memcpy(dest: dest, src: BLAKE3::hash<20>(Data: arr).data(), n: hashSize); |
2905 | }); |
2906 | break; |
2907 | case BuildIdKind::Uuid: |
2908 | if (auto ec = llvm::getRandomBytes(Buffer: buildId.get(), Size: hashSize)) |
2909 | error(msg: "entropy source failure: " + ec.message()); |
2910 | break; |
2911 | default: |
2912 | llvm_unreachable("unknown BuildIdKind" ); |
2913 | } |
2914 | for (Partition &part : partitions) |
2915 | part.buildId->writeBuildId(buf: output); |
2916 | } |
2917 | |
2918 | template void elf::writeResult<ELF32LE>(); |
2919 | template void elf::writeResult<ELF32BE>(); |
2920 | template void elf::writeResult<ELF64LE>(); |
2921 | template void elf::writeResult<ELF64BE>(); |
2922 | |