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 "BPSectionOrderer.h" |
13 | #include "CallGraphSort.h" |
14 | #include "Config.h" |
15 | #include "InputFiles.h" |
16 | #include "LinkerScript.h" |
17 | #include "MapFile.h" |
18 | #include "OutputSections.h" |
19 | #include "Relocations.h" |
20 | #include "SymbolTable.h" |
21 | #include "Symbols.h" |
22 | #include "SyntheticSections.h" |
23 | #include "Target.h" |
24 | #include "lld/Common/Arrays.h" |
25 | #include "lld/Common/CommonLinkerContext.h" |
26 | #include "lld/Common/Filesystem.h" |
27 | #include "lld/Common/Strings.h" |
28 | #include "llvm/ADT/STLExtras.h" |
29 | #include "llvm/ADT/StringMap.h" |
30 | #include "llvm/Support/BLAKE3.h" |
31 | #include "llvm/Support/Parallel.h" |
32 | #include "llvm/Support/RandomNumberGenerator.h" |
33 | #include "llvm/Support/TimeProfiler.h" |
34 | #include "llvm/Support/xxhash.h" |
35 | #include <climits> |
36 | |
37 | #define DEBUG_TYPE "lld" |
38 | |
39 | using namespace llvm; |
40 | using namespace llvm::ELF; |
41 | using namespace llvm::object; |
42 | using namespace llvm::support; |
43 | using namespace llvm::support::endian; |
44 | using namespace lld; |
45 | using namespace lld::elf; |
46 | |
47 | namespace { |
48 | // The writer writes a SymbolTable result to a file. |
49 | template <class ELFT> class Writer { |
50 | public: |
51 | LLVM_ELF_IMPORT_TYPES_ELFT(ELFT) |
52 | |
53 | Writer(Ctx &ctx) : ctx(ctx), buffer(ctx.e.outputBuffer), tc(ctx) {} |
54 | |
55 | void run(); |
56 | |
57 | private: |
58 | void addSectionSymbols(); |
59 | void sortSections(); |
60 | void resolveShfLinkOrder(); |
61 | void finalizeAddressDependentContent(); |
62 | void optimizeBasicBlockJumps(); |
63 | void sortInputSections(); |
64 | void sortOrphanSections(); |
65 | void finalizeSections(); |
66 | void checkExecuteOnly(); |
67 | void checkExecuteOnlyReport(); |
68 | void setReservedSymbolSections(); |
69 | |
70 | SmallVector<std::unique_ptr<PhdrEntry>, 0> createPhdrs(Partition &part); |
71 | void addPhdrForSection(Partition &part, unsigned shType, unsigned pType, |
72 | unsigned pFlags); |
73 | void assignFileOffsets(); |
74 | void assignFileOffsetsBinary(); |
75 | void setPhdrs(Partition &part); |
76 | void checkSections(); |
77 | void fixSectionAlignments(); |
78 | void openFile(); |
79 | void writeTrapInstr(); |
80 | void writeHeader(); |
81 | void writeSections(); |
82 | void writeSectionsBinary(); |
83 | void writeBuildId(); |
84 | |
85 | Ctx &ctx; |
86 | std::unique_ptr<FileOutputBuffer> &buffer; |
87 | // ThunkCreator holds Thunks that are used at writeTo time. |
88 | ThunkCreator tc; |
89 | |
90 | void addRelIpltSymbols(); |
91 | void addStartEndSymbols(); |
92 | void addStartStopSymbols(OutputSection &osec); |
93 | |
94 | uint64_t fileSize; |
95 | uint64_t ; |
96 | }; |
97 | } // anonymous namespace |
98 | |
99 | template <class ELFT> void elf::writeResult(Ctx &ctx) { |
100 | Writer<ELFT>(ctx).run(); |
101 | } |
102 | |
103 | static void |
104 | removeEmptyPTLoad(Ctx &ctx, SmallVector<std::unique_ptr<PhdrEntry>, 0> &phdrs) { |
105 | auto it = std::stable_partition(first: phdrs.begin(), last: phdrs.end(), pred: [&](auto &p) { |
106 | if (p->p_type != PT_LOAD) |
107 | return true; |
108 | if (!p->firstSec) |
109 | return false; |
110 | uint64_t size = p->lastSec->addr + p->lastSec->size - p->firstSec->addr; |
111 | return size != 0; |
112 | }); |
113 | |
114 | // Clear OutputSection::ptLoad for sections contained in removed |
115 | // segments. |
116 | DenseSet<PhdrEntry *> removed; |
117 | for (auto it2 = it; it2 != phdrs.end(); ++it2) |
118 | removed.insert(V: it2->get()); |
119 | for (OutputSection *sec : ctx.outputSections) |
120 | if (removed.count(V: sec->ptLoad)) |
121 | sec->ptLoad = nullptr; |
122 | phdrs.erase(CS: it, CE: phdrs.end()); |
123 | } |
124 | |
125 | void elf::copySectionsIntoPartitions(Ctx &ctx) { |
126 | SmallVector<InputSectionBase *, 0> newSections; |
127 | const size_t ehSize = ctx.ehInputSections.size(); |
128 | for (unsigned part = 2; part != ctx.partitions.size() + 1; ++part) { |
129 | for (InputSectionBase *s : ctx.inputSections) { |
130 | if (!(s->flags & SHF_ALLOC) || !s->isLive() || s->type != SHT_NOTE) |
131 | continue; |
132 | auto *copy = make<InputSection>(args&: cast<InputSection>(Val&: *s)); |
133 | copy->partition = part; |
134 | newSections.push_back(Elt: copy); |
135 | } |
136 | for (size_t i = 0; i != ehSize; ++i) { |
137 | assert(ctx.ehInputSections[i]->isLive()); |
138 | auto *copy = make<EhInputSection>(args&: *ctx.ehInputSections[i]); |
139 | copy->partition = part; |
140 | ctx.ehInputSections.push_back(Elt: copy); |
141 | } |
142 | } |
143 | |
144 | ctx.inputSections.insert(I: ctx.inputSections.end(), From: newSections.begin(), |
145 | To: newSections.end()); |
146 | } |
147 | |
148 | static Defined *addOptionalRegular(Ctx &ctx, StringRef name, SectionBase *sec, |
149 | uint64_t val, uint8_t stOther = STV_HIDDEN) { |
150 | Symbol *s = ctx.symtab->find(name); |
151 | if (!s || s->isDefined() || s->isCommon()) |
152 | return nullptr; |
153 | |
154 | ctx.synthesizedSymbols.push_back(Elt: s); |
155 | s->resolve(ctx, other: Defined{ctx, ctx.internalFile, StringRef(), STB_GLOBAL, |
156 | stOther, STT_NOTYPE, val, |
157 | /*size=*/0, sec}); |
158 | s->isUsedInRegularObj = true; |
159 | return cast<Defined>(Val: s); |
160 | } |
161 | |
162 | // The linker is expected to define some symbols depending on |
163 | // the linking result. This function defines such symbols. |
164 | void elf::addReservedSymbols(Ctx &ctx) { |
165 | if (ctx.arg.emachine == EM_MIPS) { |
166 | auto addAbsolute = [&](StringRef name) { |
167 | Symbol *sym = |
168 | ctx.symtab->addSymbol(newSym: Defined{ctx, ctx.internalFile, name, STB_GLOBAL, |
169 | STV_HIDDEN, STT_NOTYPE, 0, 0, nullptr}); |
170 | sym->isUsedInRegularObj = true; |
171 | return cast<Defined>(Val: sym); |
172 | }; |
173 | // Define _gp for MIPS. st_value of _gp symbol will be updated by Writer |
174 | // so that it points to an absolute address which by default is relative |
175 | // to GOT. Default offset is 0x7ff0. |
176 | // See "Global Data Symbols" in Chapter 6 in the following document: |
177 | // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf |
178 | ctx.sym.mipsGp = addAbsolute("_gp" ); |
179 | |
180 | // On MIPS O32 ABI, _gp_disp is a magic symbol designates offset between |
181 | // start of function and 'gp' pointer into GOT. |
182 | if (ctx.symtab->find(name: "_gp_disp" )) |
183 | ctx.sym.mipsGpDisp = addAbsolute("_gp_disp" ); |
184 | |
185 | // The __gnu_local_gp is a magic symbol equal to the current value of 'gp' |
186 | // pointer. This symbol is used in the code generated by .cpload pseudo-op |
187 | // in case of using -mno-shared option. |
188 | // https://sourceware.org/ml/binutils/2004-12/msg00094.html |
189 | if (ctx.symtab->find(name: "__gnu_local_gp" )) |
190 | ctx.sym.mipsLocalGp = addAbsolute("__gnu_local_gp" ); |
191 | } else if (ctx.arg.emachine == EM_PPC) { |
192 | // glibc *crt1.o has a undefined reference to _SDA_BASE_. Since we don't |
193 | // support Small Data Area, define it arbitrarily as 0. |
194 | addOptionalRegular(ctx, name: "_SDA_BASE_" , sec: nullptr, val: 0, stOther: STV_HIDDEN); |
195 | } else if (ctx.arg.emachine == EM_PPC64) { |
196 | addPPC64SaveRestore(ctx); |
197 | } |
198 | |
199 | // The Power Architecture 64-bit v2 ABI defines a TableOfContents (TOC) which |
200 | // combines the typical ELF GOT with the small data sections. It commonly |
201 | // includes .got .toc .sdata .sbss. The .TOC. symbol replaces both |
202 | // _GLOBAL_OFFSET_TABLE_ and _SDA_BASE_ from the 32-bit ABI. It is used to |
203 | // represent the TOC base which is offset by 0x8000 bytes from the start of |
204 | // the .got section. |
205 | // We do not allow _GLOBAL_OFFSET_TABLE_ to be defined by input objects as the |
206 | // correctness of some relocations depends on its value. |
207 | StringRef gotSymName = |
208 | (ctx.arg.emachine == EM_PPC64) ? ".TOC." : "_GLOBAL_OFFSET_TABLE_" ; |
209 | |
210 | if (Symbol *s = ctx.symtab->find(name: gotSymName)) { |
211 | if (s->isDefined()) { |
212 | ErrAlways(ctx) << s->file << " cannot redefine linker defined symbol '" |
213 | << gotSymName << "'" ; |
214 | return; |
215 | } |
216 | |
217 | uint64_t gotOff = 0; |
218 | if (ctx.arg.emachine == EM_PPC64) |
219 | gotOff = 0x8000; |
220 | |
221 | s->resolve(ctx, other: Defined{ctx, ctx.internalFile, StringRef(), STB_GLOBAL, |
222 | STV_HIDDEN, STT_NOTYPE, gotOff, /*size=*/0, |
223 | ctx.out.elfHeader.get()}); |
224 | ctx.sym.globalOffsetTable = cast<Defined>(Val: s); |
225 | } |
226 | |
227 | // __ehdr_start is the location of ELF file headers. Note that we define |
228 | // this symbol unconditionally even when using a linker script, which |
229 | // differs from the behavior implemented by GNU linker which only define |
230 | // this symbol if ELF headers are in the memory mapped segment. |
231 | addOptionalRegular(ctx, name: "__ehdr_start" , sec: ctx.out.elfHeader.get(), val: 0, |
232 | stOther: STV_HIDDEN); |
233 | |
234 | // __executable_start is not documented, but the expectation of at |
235 | // least the Android libc is that it points to the ELF header. |
236 | addOptionalRegular(ctx, name: "__executable_start" , sec: ctx.out.elfHeader.get(), val: 0, |
237 | stOther: STV_HIDDEN); |
238 | |
239 | // __dso_handle symbol is passed to cxa_finalize as a marker to identify |
240 | // each DSO. The address of the symbol doesn't matter as long as they are |
241 | // different in different DSOs, so we chose the start address of the DSO. |
242 | addOptionalRegular(ctx, name: "__dso_handle" , sec: ctx.out.elfHeader.get(), val: 0, |
243 | stOther: STV_HIDDEN); |
244 | |
245 | // If linker script do layout we do not need to create any standard symbols. |
246 | if (ctx.script->hasSectionsCommand) |
247 | return; |
248 | |
249 | auto add = [&](StringRef s, int64_t pos) { |
250 | return addOptionalRegular(ctx, name: s, sec: ctx.out.elfHeader.get(), val: pos, |
251 | stOther: STV_DEFAULT); |
252 | }; |
253 | |
254 | ctx.sym.bss = add("__bss_start" , 0); |
255 | ctx.sym.end1 = add("end" , -1); |
256 | ctx.sym.end2 = add("_end" , -1); |
257 | ctx.sym.etext1 = add("etext" , -1); |
258 | ctx.sym.etext2 = add("_etext" , -1); |
259 | ctx.sym.edata1 = add("edata" , -1); |
260 | ctx.sym.edata2 = add("_edata" , -1); |
261 | } |
262 | |
263 | static void demoteDefined(Defined &sym, DenseMap<SectionBase *, size_t> &map) { |
264 | if (map.empty()) |
265 | for (auto [i, sec] : llvm::enumerate(First: sym.file->getSections())) |
266 | map.try_emplace(Key: sec, Args&: i); |
267 | // Change WEAK to GLOBAL so that if a scanned relocation references sym, |
268 | // maybeReportUndefined will report an error. |
269 | uint8_t binding = sym.isWeak() ? uint8_t(STB_GLOBAL) : sym.binding; |
270 | Undefined(sym.file, sym.getName(), binding, sym.stOther, sym.type, |
271 | /*discardedSecIdx=*/map.lookup(Val: sym.section)) |
272 | .overwrite(sym); |
273 | // Eliminate from the symbol table, otherwise we would leave an undefined |
274 | // symbol if the symbol is unreferenced in the absence of GC. |
275 | sym.isUsedInRegularObj = false; |
276 | } |
277 | |
278 | // If all references to a DSO happen to be weak, the DSO is not added to |
279 | // DT_NEEDED. If that happens, replace ShardSymbol with Undefined to avoid |
280 | // dangling references to an unneeded DSO. Use a weak binding to avoid |
281 | // --no-allow-shlib-undefined diagnostics. Similarly, demote lazy symbols. |
282 | // |
283 | // In addition, demote symbols defined in discarded sections, so that |
284 | // references to /DISCARD/ discarded symbols will lead to errors. |
285 | static void demoteSymbolsAndComputeIsPreemptible(Ctx &ctx) { |
286 | llvm::TimeTraceScope timeScope("Demote symbols" ); |
287 | DenseMap<InputFile *, DenseMap<SectionBase *, size_t>> sectionIndexMap; |
288 | for (Symbol *sym : ctx.symtab->getSymbols()) { |
289 | if (auto *d = dyn_cast<Defined>(Val: sym)) { |
290 | if (d->section && !d->section->isLive()) |
291 | demoteDefined(sym&: *d, map&: sectionIndexMap[d->file]); |
292 | } else { |
293 | auto *s = dyn_cast<SharedSymbol>(Val: sym); |
294 | if (sym->isLazy() || (s && !cast<SharedFile>(Val: s->file)->isNeeded)) { |
295 | uint8_t binding = sym->isLazy() ? sym->binding : uint8_t(STB_WEAK); |
296 | Undefined(ctx.internalFile, sym->getName(), binding, sym->stOther, |
297 | sym->type) |
298 | .overwrite(sym&: *sym); |
299 | sym->versionId = VER_NDX_GLOBAL; |
300 | } |
301 | } |
302 | |
303 | sym->isPreemptible = (sym->isUndefined() || sym->isExported) && |
304 | computeIsPreemptible(ctx, sym: *sym); |
305 | } |
306 | } |
307 | |
308 | static OutputSection *findSection(Ctx &ctx, StringRef name, |
309 | unsigned partition = 1) { |
310 | for (SectionCommand *cmd : ctx.script->sectionCommands) |
311 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) |
312 | if (osd->osec.name == name && osd->osec.partition == partition) |
313 | return &osd->osec; |
314 | return nullptr; |
315 | } |
316 | |
317 | // The main function of the writer. |
318 | template <class ELFT> void Writer<ELFT>::run() { |
319 | // Now that we have a complete set of output sections. This function |
320 | // completes section contents. For example, we need to add strings |
321 | // to the string table, and add entries to .got and .plt. |
322 | // finalizeSections does that. |
323 | finalizeSections(); |
324 | checkExecuteOnly(); |
325 | checkExecuteOnlyReport(); |
326 | |
327 | // If --compressed-debug-sections is specified, compress .debug_* sections. |
328 | // Do it right now because it changes the size of output sections. |
329 | for (OutputSection *sec : ctx.outputSections) |
330 | sec->maybeCompress<ELFT>(ctx); |
331 | |
332 | if (ctx.script->hasSectionsCommand) |
333 | ctx.script->allocateHeaders(phdrs&: ctx.mainPart->phdrs); |
334 | |
335 | // Remove empty PT_LOAD to avoid causing the dynamic linker to try to mmap a |
336 | // 0 sized region. This has to be done late since only after assignAddresses |
337 | // we know the size of the sections. |
338 | for (Partition &part : ctx.partitions) |
339 | removeEmptyPTLoad(ctx, phdrs&: part.phdrs); |
340 | |
341 | if (!ctx.arg.oFormatBinary) |
342 | assignFileOffsets(); |
343 | else |
344 | assignFileOffsetsBinary(); |
345 | |
346 | for (Partition &part : ctx.partitions) |
347 | setPhdrs(part); |
348 | |
349 | // Handle --print-map(-M)/--Map and --cref. Dump them before checkSections() |
350 | // because the files may be useful in case checkSections() or openFile() |
351 | // fails, for example, due to an erroneous file size. |
352 | writeMapAndCref(ctx); |
353 | |
354 | // Handle --print-memory-usage option. |
355 | if (ctx.arg.printMemoryUsage) |
356 | ctx.script->printMemoryUsage(os&: ctx.e.outs()); |
357 | |
358 | if (ctx.arg.checkSections) |
359 | checkSections(); |
360 | |
361 | // It does not make sense try to open the file if we have error already. |
362 | if (errCount(ctx)) |
363 | return; |
364 | |
365 | { |
366 | llvm::TimeTraceScope timeScope("Write output file" ); |
367 | // Write the result down to a file. |
368 | openFile(); |
369 | if (errCount(ctx)) |
370 | return; |
371 | |
372 | if (!ctx.arg.oFormatBinary) { |
373 | if (ctx.arg.zSeparate != SeparateSegmentKind::None) |
374 | writeTrapInstr(); |
375 | writeHeader(); |
376 | writeSections(); |
377 | } else { |
378 | writeSectionsBinary(); |
379 | } |
380 | |
381 | // Backfill .note.gnu.build-id section content. This is done at last |
382 | // because the content is usually a hash value of the entire output file. |
383 | writeBuildId(); |
384 | if (errCount(ctx)) |
385 | return; |
386 | |
387 | if (!ctx.e.disableOutput) { |
388 | if (auto e = buffer->commit()) |
389 | Err(ctx) << "failed to write output '" << buffer->getPath() |
390 | << "': " << std::move(e); |
391 | } |
392 | |
393 | if (!ctx.arg.cmseOutputLib.empty()) |
394 | writeARMCmseImportLib<ELFT>(ctx); |
395 | } |
396 | } |
397 | |
398 | template <class ELFT, class RelTy> |
399 | static void markUsedLocalSymbolsImpl(ObjFile<ELFT> *file, |
400 | llvm::ArrayRef<RelTy> rels) { |
401 | for (const RelTy &rel : rels) { |
402 | Symbol &sym = file->getRelocTargetSym(rel); |
403 | if (sym.isLocal()) |
404 | sym.used = true; |
405 | } |
406 | } |
407 | |
408 | // The function ensures that the "used" field of local symbols reflects the fact |
409 | // that the symbol is used in a relocation from a live section. |
410 | template <class ELFT> static void markUsedLocalSymbols(Ctx &ctx) { |
411 | // With --gc-sections, the field is already filled. |
412 | // See MarkLive<ELFT>::resolveReloc(). |
413 | if (ctx.arg.gcSections) |
414 | return; |
415 | for (ELFFileBase *file : ctx.objectFiles) { |
416 | ObjFile<ELFT> *f = cast<ObjFile<ELFT>>(file); |
417 | for (InputSectionBase *s : f->getSections()) { |
418 | InputSection *isec = dyn_cast_or_null<InputSection>(Val: s); |
419 | if (!isec) |
420 | continue; |
421 | if (isec->type == SHT_REL) { |
422 | markUsedLocalSymbolsImpl(f, isec->getDataAs<typename ELFT::Rel>()); |
423 | } else if (isec->type == SHT_RELA) { |
424 | markUsedLocalSymbolsImpl(f, isec->getDataAs<typename ELFT::Rela>()); |
425 | } else if (isec->type == SHT_CREL) { |
426 | // The is64=true variant also works with ELF32 since only the r_symidx |
427 | // member is used. |
428 | for (Elf_Crel_Impl<true> r : RelocsCrel<true>(isec->content_)) { |
429 | Symbol &sym = file->getSymbol(symbolIndex: r.r_symidx); |
430 | if (sym.isLocal()) |
431 | sym.used = true; |
432 | } |
433 | } |
434 | } |
435 | } |
436 | } |
437 | |
438 | static bool shouldKeepInSymtab(Ctx &ctx, const Defined &sym) { |
439 | if (sym.isSection()) |
440 | return false; |
441 | |
442 | // If --emit-reloc or -r is given, preserve symbols referenced by relocations |
443 | // from live sections. |
444 | if (sym.used && ctx.arg.copyRelocs) |
445 | return true; |
446 | |
447 | // Exclude local symbols pointing to .ARM.exidx sections. |
448 | // They are probably mapping symbols "$d", which are optional for these |
449 | // sections. After merging the .ARM.exidx sections, some of these symbols |
450 | // may become dangling. The easiest way to avoid the issue is not to add |
451 | // them to the symbol table from the beginning. |
452 | if (ctx.arg.emachine == EM_ARM && sym.section && |
453 | sym.section->type == SHT_ARM_EXIDX) |
454 | return false; |
455 | |
456 | if (ctx.arg.discard == DiscardPolicy::None) |
457 | return true; |
458 | if (ctx.arg.discard == DiscardPolicy::All) |
459 | return false; |
460 | |
461 | // In ELF assembly .L symbols are normally discarded by the assembler. |
462 | // If the assembler fails to do so, the linker discards them if |
463 | // * --discard-locals is used. |
464 | // * The symbol is in a SHF_MERGE section, which is normally the reason for |
465 | // the assembler keeping the .L symbol. |
466 | if (sym.getName().starts_with(Prefix: ".L" ) && |
467 | (ctx.arg.discard == DiscardPolicy::Locals || |
468 | (sym.section && (sym.section->flags & SHF_MERGE)))) |
469 | return false; |
470 | return true; |
471 | } |
472 | |
473 | bool elf::includeInSymtab(Ctx &ctx, const Symbol &b) { |
474 | if (auto *d = dyn_cast<Defined>(Val: &b)) { |
475 | // Always include absolute symbols. |
476 | SectionBase *sec = d->section; |
477 | if (!sec) |
478 | return true; |
479 | assert(sec->isLive()); |
480 | |
481 | if (auto *s = dyn_cast<MergeInputSection>(Val: sec)) |
482 | return s->getSectionPiece(offset: d->value).live; |
483 | return true; |
484 | } |
485 | return b.used || !ctx.arg.gcSections; |
486 | } |
487 | |
488 | // Scan local symbols to: |
489 | // |
490 | // - demote symbols defined relative to /DISCARD/ discarded input sections so |
491 | // that relocations referencing them will lead to errors. |
492 | // - copy eligible symbols to .symTab |
493 | static void demoteAndCopyLocalSymbols(Ctx &ctx) { |
494 | llvm::TimeTraceScope timeScope("Add local symbols" ); |
495 | for (ELFFileBase *file : ctx.objectFiles) { |
496 | DenseMap<SectionBase *, size_t> sectionIndexMap; |
497 | for (Symbol *b : file->getLocalSymbols()) { |
498 | assert(b->isLocal() && "should have been caught in initializeSymbols()" ); |
499 | auto *dr = dyn_cast<Defined>(Val: b); |
500 | if (!dr) |
501 | continue; |
502 | |
503 | if (dr->section && !dr->section->isLive()) |
504 | demoteDefined(sym&: *dr, map&: sectionIndexMap); |
505 | else if (ctx.in.symTab && includeInSymtab(ctx, b: *b) && |
506 | shouldKeepInSymtab(ctx, sym: *dr)) |
507 | ctx.in.symTab->addSymbol(sym: b); |
508 | } |
509 | } |
510 | } |
511 | |
512 | // Create a section symbol for each output section so that we can represent |
513 | // relocations that point to the section. If we know that no relocation is |
514 | // referring to a section (that happens if the section is a synthetic one), we |
515 | // don't create a section symbol for that section. |
516 | template <class ELFT> void Writer<ELFT>::addSectionSymbols() { |
517 | for (SectionCommand *cmd : ctx.script->sectionCommands) { |
518 | auto *osd = dyn_cast<OutputDesc>(Val: cmd); |
519 | if (!osd) |
520 | continue; |
521 | OutputSection &osec = osd->osec; |
522 | InputSectionBase *isec = nullptr; |
523 | // Iterate over all input sections and add a STT_SECTION symbol if any input |
524 | // section may be a relocation target. |
525 | for (SectionCommand *cmd : osec.commands) { |
526 | auto *isd = dyn_cast<InputSectionDescription>(Val: cmd); |
527 | if (!isd) |
528 | continue; |
529 | for (InputSectionBase *s : isd->sections) { |
530 | // Relocations are not using REL[A] section symbols. |
531 | if (isStaticRelSecType(type: s->type)) |
532 | continue; |
533 | |
534 | // Unlike other synthetic sections, mergeable output sections contain |
535 | // data copied from input sections, and there may be a relocation |
536 | // pointing to its contents if -r or --emit-reloc is given. |
537 | if (isa<SyntheticSection>(Val: s) && !(s->flags & SHF_MERGE)) |
538 | continue; |
539 | |
540 | isec = s; |
541 | break; |
542 | } |
543 | } |
544 | if (!isec) |
545 | continue; |
546 | |
547 | // Set the symbol to be relative to the output section so that its st_value |
548 | // equals the output section address. Note, there may be a gap between the |
549 | // start of the output section and isec. |
550 | ctx.in.symTab->addSymbol(sym: makeDefined(args&: ctx, args&: isec->file, args: "" , args: STB_LOCAL, |
551 | /*stOther=*/args: 0, args: STT_SECTION, |
552 | /*value=*/args: 0, /*size=*/args: 0, args: &osec)); |
553 | } |
554 | } |
555 | |
556 | // Today's loaders have a feature to make segments read-only after |
557 | // processing dynamic relocations to enhance security. PT_GNU_RELRO |
558 | // is defined for that. |
559 | // |
560 | // This function returns true if a section needs to be put into a |
561 | // PT_GNU_RELRO segment. |
562 | static bool isRelroSection(Ctx &ctx, const OutputSection *sec) { |
563 | if (!ctx.arg.zRelro) |
564 | return false; |
565 | if (sec->relro) |
566 | return true; |
567 | |
568 | uint64_t flags = sec->flags; |
569 | |
570 | // Non-allocatable or non-writable sections don't need RELRO because |
571 | // they are not writable or not even mapped to memory in the first place. |
572 | // RELRO is for sections that are essentially read-only but need to |
573 | // be writable only at process startup to allow dynamic linker to |
574 | // apply relocations. |
575 | if (!(flags & SHF_ALLOC) || !(flags & SHF_WRITE)) |
576 | return false; |
577 | |
578 | // Once initialized, TLS data segments are used as data templates |
579 | // for a thread-local storage. For each new thread, runtime |
580 | // allocates memory for a TLS and copy templates there. No thread |
581 | // are supposed to use templates directly. Thus, it can be in RELRO. |
582 | if (flags & SHF_TLS) |
583 | return true; |
584 | |
585 | // .init_array, .preinit_array and .fini_array contain pointers to |
586 | // functions that are executed on process startup or exit. These |
587 | // pointers are set by the static linker, and they are not expected |
588 | // to change at runtime. But if you are an attacker, you could do |
589 | // interesting things by manipulating pointers in .fini_array, for |
590 | // example. So they are put into RELRO. |
591 | uint32_t type = sec->type; |
592 | if (type == SHT_INIT_ARRAY || type == SHT_FINI_ARRAY || |
593 | type == SHT_PREINIT_ARRAY) |
594 | return true; |
595 | |
596 | // .got contains pointers to external symbols. They are resolved by |
597 | // the dynamic linker when a module is loaded into memory, and after |
598 | // that they are not expected to change. So, it can be in RELRO. |
599 | if (ctx.in.got && sec == ctx.in.got->getParent()) |
600 | return true; |
601 | |
602 | // .toc is a GOT-ish section for PowerPC64. Their contents are accessed |
603 | // through r2 register, which is reserved for that purpose. Since r2 is used |
604 | // for accessing .got as well, .got and .toc need to be close enough in the |
605 | // virtual address space. Usually, .toc comes just after .got. Since we place |
606 | // .got into RELRO, .toc needs to be placed into RELRO too. |
607 | if (sec->name == ".toc" ) |
608 | return true; |
609 | |
610 | // .got.plt contains pointers to external function symbols. They are |
611 | // by default resolved lazily, so we usually cannot put it into RELRO. |
612 | // However, if "-z now" is given, the lazy symbol resolution is |
613 | // disabled, which enables us to put it into RELRO. |
614 | if (sec == ctx.in.gotPlt->getParent()) |
615 | return ctx.arg.zNow; |
616 | |
617 | if (ctx.in.relroPadding && sec == ctx.in.relroPadding->getParent()) |
618 | return true; |
619 | |
620 | // .dynamic section contains data for the dynamic linker, and |
621 | // there's no need to write to it at runtime, so it's better to put |
622 | // it into RELRO. |
623 | if (sec->name == ".dynamic" ) |
624 | return true; |
625 | |
626 | // Sections with some special names are put into RELRO. This is a |
627 | // bit unfortunate because section names shouldn't be significant in |
628 | // ELF in spirit. But in reality many linker features depend on |
629 | // magic section names. |
630 | StringRef s = sec->name; |
631 | |
632 | bool abiAgnostic = s == ".data.rel.ro" || s == ".bss.rel.ro" || |
633 | s == ".ctors" || s == ".dtors" || s == ".jcr" || |
634 | s == ".eh_frame" || s == ".fini_array" || |
635 | s == ".init_array" || s == ".preinit_array" ; |
636 | |
637 | bool abiSpecific = |
638 | ctx.arg.osabi == ELFOSABI_OPENBSD && s == ".openbsd.randomdata" ; |
639 | |
640 | return abiAgnostic || abiSpecific; |
641 | } |
642 | |
643 | // We compute a rank for each section. The rank indicates where the |
644 | // section should be placed in the file. Instead of using simple |
645 | // numbers (0,1,2...), we use a series of flags. One for each decision |
646 | // point when placing the section. |
647 | // Using flags has two key properties: |
648 | // * It is easy to check if a give branch was taken. |
649 | // * It is easy two see how similar two ranks are (see getRankProximity). |
650 | enum RankFlags { |
651 | RF_NOT_ADDR_SET = 1 << 27, |
652 | RF_NOT_ALLOC = 1 << 26, |
653 | RF_PARTITION = 1 << 18, // Partition number (8 bits) |
654 | RF_LARGE_EXEC_WRITE = 1 << 16, |
655 | RF_LARGE_ALT = 1 << 15, |
656 | RF_WRITE = 1 << 14, |
657 | RF_EXEC_WRITE = 1 << 13, |
658 | RF_EXEC = 1 << 12, |
659 | RF_RODATA = 1 << 11, |
660 | RF_LARGE_EXEC = 1 << 10, |
661 | RF_LARGE = 1 << 9, |
662 | RF_NOT_RELRO = 1 << 8, |
663 | RF_NOT_TLS = 1 << 7, |
664 | RF_BSS = 1 << 6, |
665 | }; |
666 | |
667 | unsigned elf::getSectionRank(Ctx &ctx, OutputSection &osec) { |
668 | unsigned rank = osec.partition * RF_PARTITION; |
669 | |
670 | // We want to put section specified by -T option first, so we |
671 | // can start assigning VA starting from them later. |
672 | if (ctx.arg.sectionStartMap.count(Key: osec.name)) |
673 | return rank; |
674 | rank |= RF_NOT_ADDR_SET; |
675 | |
676 | // Allocatable sections go first to reduce the total PT_LOAD size and |
677 | // so debug info doesn't change addresses in actual code. |
678 | if (!(osec.flags & SHF_ALLOC)) |
679 | return rank | RF_NOT_ALLOC; |
680 | |
681 | // Sort sections based on their access permission in the following |
682 | // order: R, RX, RXW, RW(RELRO), RW(non-RELRO). |
683 | // |
684 | // Read-only sections come first such that they go in the PT_LOAD covering the |
685 | // program headers at the start of the file. |
686 | // |
687 | // The layout for writable sections is PT_LOAD(PT_GNU_RELRO(.data.rel.ro |
688 | // .bss.rel.ro) | .data .bss), where | marks where page alignment happens. |
689 | // An alternative ordering is PT_LOAD(.data | PT_GNU_RELRO( .data.rel.ro |
690 | // .bss.rel.ro) | .bss), but it may waste more bytes due to 2 alignment |
691 | // places. |
692 | bool isExec = osec.flags & SHF_EXECINSTR; |
693 | bool isWrite = osec.flags & SHF_WRITE; |
694 | bool isLarge = osec.flags & SHF_X86_64_LARGE && ctx.arg.emachine == EM_X86_64; |
695 | |
696 | if (!isWrite && !isExec) { |
697 | // Among PROGBITS sections, place .lrodata further from .text. |
698 | // For -z lrodata-after-bss, place .lrodata after .lbss like GNU ld. This |
699 | // layout has one extra PT_LOAD, but alleviates relocation overflow |
700 | // pressure for absolute relocations referencing small data from -fno-pic |
701 | // relocatable files. |
702 | if (isLarge) |
703 | rank |= ctx.arg.zLrodataAfterBss ? RF_LARGE_ALT : 0; |
704 | else |
705 | rank |= ctx.arg.zLrodataAfterBss ? 0 : RF_LARGE; |
706 | |
707 | if (osec.type == SHT_LLVM_PART_EHDR) |
708 | ; |
709 | else if (osec.type == SHT_LLVM_PART_PHDR) |
710 | rank |= 1; |
711 | else if (osec.name == ".interp" ) |
712 | rank |= 2; |
713 | // Put .note sections at the beginning so that they are likely to be |
714 | // included in a truncate core file. In particular, .note.gnu.build-id, if |
715 | // available, can identify the object file. |
716 | else if (osec.type == SHT_NOTE) |
717 | rank |= 3; |
718 | // Make PROGBITS sections (e.g .rodata .eh_frame) closer to .text to |
719 | // alleviate relocation overflow pressure. Large special sections such as |
720 | // .dynstr and .dynsym can be away from .text. |
721 | else if (osec.type != SHT_PROGBITS) |
722 | rank |= 4; |
723 | else |
724 | rank |= RF_RODATA; |
725 | } else if (isExec) { |
726 | // Place readonly .ltext before .lrodata and writable .ltext after .lbss to |
727 | // keep writable and readonly segments separate. |
728 | if (isLarge) { |
729 | rank |= isWrite ? RF_LARGE_EXEC_WRITE : RF_LARGE_EXEC; |
730 | } else { |
731 | rank |= isWrite ? RF_EXEC_WRITE : RF_EXEC; |
732 | } |
733 | } else { |
734 | rank |= RF_WRITE; |
735 | // The TLS initialization block needs to be a single contiguous block. Place |
736 | // TLS sections directly before the other RELRO sections. |
737 | if (!(osec.flags & SHF_TLS)) |
738 | rank |= RF_NOT_TLS; |
739 | if (isRelroSection(ctx, sec: &osec)) |
740 | osec.relro = true; |
741 | else |
742 | rank |= RF_NOT_RELRO; |
743 | // Place .ldata and .lbss after .bss. Making .bss closer to .text |
744 | // alleviates relocation overflow pressure. |
745 | // For -z lrodata-after-bss, place .lbss/.lrodata/.ldata after .bss. |
746 | // .bss/.lbss being adjacent reuses the NOBITS size optimization. |
747 | if (isLarge) { |
748 | rank |= ctx.arg.zLrodataAfterBss |
749 | ? (osec.type == SHT_NOBITS ? 1 : RF_LARGE_ALT) |
750 | : RF_LARGE; |
751 | } |
752 | } |
753 | |
754 | // Within TLS sections, or within other RelRo sections, or within non-RelRo |
755 | // sections, place non-NOBITS sections first. |
756 | if (osec.type == SHT_NOBITS) |
757 | rank |= RF_BSS; |
758 | |
759 | // Some architectures have additional ordering restrictions for sections |
760 | // within the same PT_LOAD. |
761 | if (ctx.arg.emachine == EM_PPC64) { |
762 | // PPC64 has a number of special SHT_PROGBITS+SHF_ALLOC+SHF_WRITE sections |
763 | // that we would like to make sure appear is a specific order to maximize |
764 | // their coverage by a single signed 16-bit offset from the TOC base |
765 | // pointer. |
766 | StringRef name = osec.name; |
767 | if (name == ".got" ) |
768 | rank |= 1; |
769 | else if (name == ".toc" ) |
770 | rank |= 2; |
771 | } |
772 | |
773 | if (ctx.arg.emachine == EM_MIPS) { |
774 | if (osec.name != ".got" ) |
775 | rank |= 1; |
776 | // All sections with SHF_MIPS_GPREL flag should be grouped together |
777 | // because data in these sections is addressable with a gp relative address. |
778 | if (osec.flags & SHF_MIPS_GPREL) |
779 | rank |= 2; |
780 | } |
781 | |
782 | if (ctx.arg.emachine == EM_RISCV) { |
783 | // .sdata and .sbss are placed closer to make GP relaxation more profitable |
784 | // and match GNU ld. |
785 | StringRef name = osec.name; |
786 | if (name == ".sdata" || (osec.type == SHT_NOBITS && name != ".sbss" )) |
787 | rank |= 1; |
788 | } |
789 | |
790 | return rank; |
791 | } |
792 | |
793 | static bool compareSections(Ctx &ctx, const SectionCommand *aCmd, |
794 | const SectionCommand *bCmd) { |
795 | const OutputSection *a = &cast<OutputDesc>(Val: aCmd)->osec; |
796 | const OutputSection *b = &cast<OutputDesc>(Val: bCmd)->osec; |
797 | |
798 | if (a->sortRank != b->sortRank) |
799 | return a->sortRank < b->sortRank; |
800 | |
801 | if (!(a->sortRank & RF_NOT_ADDR_SET)) |
802 | return ctx.arg.sectionStartMap.lookup(Key: a->name) < |
803 | ctx.arg.sectionStartMap.lookup(Key: b->name); |
804 | return false; |
805 | } |
806 | |
807 | void PhdrEntry::add(OutputSection *sec) { |
808 | lastSec = sec; |
809 | if (!firstSec) |
810 | firstSec = sec; |
811 | p_align = std::max(a: p_align, b: sec->addralign); |
812 | if (p_type == PT_LOAD) |
813 | sec->ptLoad = this; |
814 | } |
815 | |
816 | // A statically linked position-dependent executable should only contain |
817 | // IRELATIVE relocations and no other dynamic relocations. Encapsulation symbols |
818 | // __rel[a]_iplt_{start,end} will be defined for .rel[a].dyn, to be |
819 | // processed by the libc runtime. Other executables or DSOs use dynamic tags |
820 | // instead. |
821 | template <class ELFT> void Writer<ELFT>::addRelIpltSymbols() { |
822 | if (ctx.arg.isPic) |
823 | return; |
824 | |
825 | // __rela_iplt_{start,end} are initially defined relative to dummy section 0. |
826 | // We'll override ctx.out.elfHeader with relaDyn later when we are sure that |
827 | // .rela.dyn will be present in the output. |
828 | std::string name = ctx.arg.isRela ? "__rela_iplt_start" : "__rel_iplt_start" ; |
829 | ctx.sym.relaIpltStart = |
830 | addOptionalRegular(ctx, name, sec: ctx.out.elfHeader.get(), val: 0, stOther: STV_HIDDEN); |
831 | name.replace(pos: name.size() - 5, n1: 5, s: "end" ); |
832 | ctx.sym.relaIpltEnd = |
833 | addOptionalRegular(ctx, name, sec: ctx.out.elfHeader.get(), val: 0, stOther: STV_HIDDEN); |
834 | } |
835 | |
836 | // This function generates assignments for predefined symbols (e.g. _end or |
837 | // _etext) and inserts them into the commands sequence to be processed at the |
838 | // appropriate time. This ensures that the value is going to be correct by the |
839 | // time any references to these symbols are processed and is equivalent to |
840 | // defining these symbols explicitly in the linker script. |
841 | template <class ELFT> void Writer<ELFT>::setReservedSymbolSections() { |
842 | if (ctx.sym.globalOffsetTable) { |
843 | // The _GLOBAL_OFFSET_TABLE_ symbol is defined by target convention usually |
844 | // to the start of the .got or .got.plt section. |
845 | InputSection *sec = ctx.in.gotPlt.get(); |
846 | if (!ctx.target->gotBaseSymInGotPlt) |
847 | sec = ctx.in.mipsGot ? cast<InputSection>(Val: ctx.in.mipsGot.get()) |
848 | : cast<InputSection>(Val: ctx.in.got.get()); |
849 | ctx.sym.globalOffsetTable->section = sec; |
850 | } |
851 | |
852 | // .rela_iplt_{start,end} mark the start and the end of the section containing |
853 | // IRELATIVE relocations. |
854 | if (ctx.sym.relaIpltStart) { |
855 | auto &dyn = getIRelativeSection(ctx); |
856 | if (dyn.isNeeded()) { |
857 | ctx.sym.relaIpltStart->section = &dyn; |
858 | ctx.sym.relaIpltEnd->section = &dyn; |
859 | ctx.sym.relaIpltEnd->value = dyn.getSize(); |
860 | } |
861 | } |
862 | |
863 | PhdrEntry *last = nullptr; |
864 | OutputSection *lastRO = nullptr; |
865 | auto isLarge = [&ctx = ctx](OutputSection *osec) { |
866 | return ctx.arg.emachine == EM_X86_64 && osec->flags & SHF_X86_64_LARGE; |
867 | }; |
868 | for (Partition &part : ctx.partitions) { |
869 | for (auto &p : part.phdrs) { |
870 | if (p->p_type != PT_LOAD) |
871 | continue; |
872 | last = p.get(); |
873 | if (!(p->p_flags & PF_W) && p->lastSec && !isLarge(p->lastSec)) |
874 | lastRO = p->lastSec; |
875 | } |
876 | } |
877 | |
878 | if (lastRO) { |
879 | // _etext is the first location after the last read-only loadable segment |
880 | // that does not contain large sections. |
881 | if (ctx.sym.etext1) |
882 | ctx.sym.etext1->section = lastRO; |
883 | if (ctx.sym.etext2) |
884 | ctx.sym.etext2->section = lastRO; |
885 | } |
886 | |
887 | if (last) { |
888 | // _edata points to the end of the last non-large mapped initialized |
889 | // section. |
890 | OutputSection *edata = nullptr; |
891 | for (OutputSection *os : ctx.outputSections) { |
892 | if (os->type != SHT_NOBITS && !isLarge(os)) |
893 | edata = os; |
894 | if (os == last->lastSec) |
895 | break; |
896 | } |
897 | |
898 | if (ctx.sym.edata1) |
899 | ctx.sym.edata1->section = edata; |
900 | if (ctx.sym.edata2) |
901 | ctx.sym.edata2->section = edata; |
902 | |
903 | // _end is the first location after the uninitialized data region. |
904 | if (ctx.sym.end1) |
905 | ctx.sym.end1->section = last->lastSec; |
906 | if (ctx.sym.end2) |
907 | ctx.sym.end2->section = last->lastSec; |
908 | } |
909 | |
910 | if (ctx.sym.bss) { |
911 | // On RISC-V, set __bss_start to the start of .sbss if present. |
912 | OutputSection *sbss = |
913 | ctx.arg.emachine == EM_RISCV ? findSection(ctx, name: ".sbss" ) : nullptr; |
914 | ctx.sym.bss->section = sbss ? sbss : findSection(ctx, name: ".bss" ); |
915 | } |
916 | |
917 | // Setup MIPS _gp_disp/__gnu_local_gp symbols which should |
918 | // be equal to the _gp symbol's value. |
919 | if (ctx.sym.mipsGp) { |
920 | // Find GP-relative section with the lowest address |
921 | // and use this address to calculate default _gp value. |
922 | for (OutputSection *os : ctx.outputSections) { |
923 | if (os->flags & SHF_MIPS_GPREL) { |
924 | ctx.sym.mipsGp->section = os; |
925 | ctx.sym.mipsGp->value = 0x7ff0; |
926 | break; |
927 | } |
928 | } |
929 | } |
930 | } |
931 | |
932 | // We want to find how similar two ranks are. |
933 | // The more branches in getSectionRank that match, the more similar they are. |
934 | // Since each branch corresponds to a bit flag, we can just use |
935 | // countLeadingZeros. |
936 | static int getRankProximity(OutputSection *a, SectionCommand *b) { |
937 | auto *osd = dyn_cast<OutputDesc>(Val: b); |
938 | return (osd && osd->osec.hasInputSections) |
939 | ? llvm::countl_zero(Val: a->sortRank ^ osd->osec.sortRank) |
940 | : -1; |
941 | } |
942 | |
943 | // When placing orphan sections, we want to place them after symbol assignments |
944 | // so that an orphan after |
945 | // begin_foo = .; |
946 | // foo : { *(foo) } |
947 | // end_foo = .; |
948 | // doesn't break the intended meaning of the begin/end symbols. |
949 | // We don't want to go over sections since findOrphanPos is the |
950 | // one in charge of deciding the order of the sections. |
951 | // We don't want to go over changes to '.', since doing so in |
952 | // rx_sec : { *(rx_sec) } |
953 | // . = ALIGN(0x1000); |
954 | // /* The RW PT_LOAD starts here*/ |
955 | // rw_sec : { *(rw_sec) } |
956 | // would mean that the RW PT_LOAD would become unaligned. |
957 | static bool shouldSkip(SectionCommand *cmd) { |
958 | if (auto *assign = dyn_cast<SymbolAssignment>(Val: cmd)) |
959 | return assign->name != "." ; |
960 | return false; |
961 | } |
962 | |
963 | // We want to place orphan sections so that they share as much |
964 | // characteristics with their neighbors as possible. For example, if |
965 | // both are rw, or both are tls. |
966 | static SmallVectorImpl<SectionCommand *>::iterator |
967 | findOrphanPos(Ctx &ctx, SmallVectorImpl<SectionCommand *>::iterator b, |
968 | SmallVectorImpl<SectionCommand *>::iterator e) { |
969 | // Place non-alloc orphan sections at the end. This matches how we assign file |
970 | // offsets to non-alloc sections. |
971 | OutputSection *sec = &cast<OutputDesc>(Val: *e)->osec; |
972 | if (!(sec->flags & SHF_ALLOC)) |
973 | return e; |
974 | |
975 | // As a special case, place .relro_padding before the SymbolAssignment using |
976 | // DATA_SEGMENT_RELRO_END, if present. |
977 | if (ctx.in.relroPadding && sec == ctx.in.relroPadding->getParent()) { |
978 | auto i = std::find_if(first: b, last: e, pred: [=](SectionCommand *a) { |
979 | if (auto *assign = dyn_cast<SymbolAssignment>(Val: a)) |
980 | return assign->dataSegmentRelroEnd; |
981 | return false; |
982 | }); |
983 | if (i != e) |
984 | return i; |
985 | } |
986 | |
987 | // Find the most similar output section as the anchor. Rank Proximity is a |
988 | // value in the range [-1, 32] where [0, 32] indicates potential anchors (0: |
989 | // least similar; 32: identical). -1 means not an anchor. |
990 | // |
991 | // In the event of proximity ties, we select the first or last section |
992 | // depending on whether the orphan's rank is smaller. |
993 | int maxP = 0; |
994 | auto i = e; |
995 | for (auto j = b; j != e; ++j) { |
996 | int p = getRankProximity(a: sec, b: *j); |
997 | if (p > maxP || |
998 | (p == maxP && cast<OutputDesc>(Val: *j)->osec.sortRank <= sec->sortRank)) { |
999 | maxP = p; |
1000 | i = j; |
1001 | } |
1002 | } |
1003 | if (i == e) |
1004 | return e; |
1005 | |
1006 | auto isOutputSecWithInputSections = [](SectionCommand *cmd) { |
1007 | auto *osd = dyn_cast<OutputDesc>(Val: cmd); |
1008 | return osd && osd->osec.hasInputSections; |
1009 | }; |
1010 | |
1011 | // Then, scan backward or forward through the script for a suitable insertion |
1012 | // point. If i's rank is larger, the orphan section can be placed before i. |
1013 | // |
1014 | // However, don't do this if custom program headers are defined. Otherwise, |
1015 | // adding the orphan to a previous segment can change its flags, for example, |
1016 | // making a read-only segment writable. If memory regions are defined, an |
1017 | // orphan section should continue the same region as the found section to |
1018 | // better resemble the behavior of GNU ld. |
1019 | bool mustAfter = |
1020 | ctx.script->hasPhdrsCommands() || !ctx.script->memoryRegions.empty(); |
1021 | if (cast<OutputDesc>(Val: *i)->osec.sortRank <= sec->sortRank || mustAfter) { |
1022 | for (auto j = ++i; j != e; ++j) { |
1023 | if (!isOutputSecWithInputSections(*j)) |
1024 | continue; |
1025 | if (getRankProximity(a: sec, b: *j) != maxP) |
1026 | break; |
1027 | i = j + 1; |
1028 | } |
1029 | } else { |
1030 | for (; i != b; --i) |
1031 | if (isOutputSecWithInputSections(i[-1])) |
1032 | break; |
1033 | } |
1034 | |
1035 | // As a special case, if the orphan section is the last section, put |
1036 | // it at the very end, past any other commands. |
1037 | // This matches bfd's behavior and is convenient when the linker script fully |
1038 | // specifies the start of the file, but doesn't care about the end (the non |
1039 | // alloc sections for example). |
1040 | if (std::none_of(first: i, last: e, pred: isOutputSecWithInputSections)) |
1041 | return e; |
1042 | |
1043 | while (i != e && shouldSkip(cmd: *i)) |
1044 | ++i; |
1045 | return i; |
1046 | } |
1047 | |
1048 | // Adds random priorities to sections not already in the map. |
1049 | static void maybeShuffle(Ctx &ctx, |
1050 | DenseMap<const InputSectionBase *, int> &order) { |
1051 | if (ctx.arg.shuffleSections.empty()) |
1052 | return; |
1053 | |
1054 | SmallVector<InputSectionBase *, 0> matched, sections = ctx.inputSections; |
1055 | matched.reserve(N: sections.size()); |
1056 | for (const auto &patAndSeed : ctx.arg.shuffleSections) { |
1057 | matched.clear(); |
1058 | for (InputSectionBase *sec : sections) |
1059 | if (patAndSeed.first.match(S: sec->name)) |
1060 | matched.push_back(Elt: sec); |
1061 | const uint32_t seed = patAndSeed.second; |
1062 | if (seed == UINT32_MAX) { |
1063 | // If --shuffle-sections <section-glob>=-1, reverse the section order. The |
1064 | // section order is stable even if the number of sections changes. This is |
1065 | // useful to catch issues like static initialization order fiasco |
1066 | // reliably. |
1067 | std::reverse(first: matched.begin(), last: matched.end()); |
1068 | } else { |
1069 | std::mt19937 g(seed ? seed : std::random_device()()); |
1070 | llvm::shuffle(first: matched.begin(), last: matched.end(), g); |
1071 | } |
1072 | size_t i = 0; |
1073 | for (InputSectionBase *&sec : sections) |
1074 | if (patAndSeed.first.match(S: sec->name)) |
1075 | sec = matched[i++]; |
1076 | } |
1077 | |
1078 | // Existing priorities are < 0, so use priorities >= 0 for the missing |
1079 | // sections. |
1080 | int prio = 0; |
1081 | for (InputSectionBase *sec : sections) { |
1082 | if (order.try_emplace(Key: sec, Args&: prio).second) |
1083 | ++prio; |
1084 | } |
1085 | } |
1086 | |
1087 | // Return section order within an InputSectionDescription. |
1088 | // If both --symbol-ordering-file and call graph profile are present, the order |
1089 | // file takes precedence, but the call graph profile is still used for symbols |
1090 | // that don't appear in the order file. |
1091 | static DenseMap<const InputSectionBase *, int> buildSectionOrder(Ctx &ctx) { |
1092 | DenseMap<const InputSectionBase *, int> sectionOrder; |
1093 | if (ctx.arg.bpStartupFunctionSort || ctx.arg.bpFunctionOrderForCompression || |
1094 | ctx.arg.bpDataOrderForCompression) { |
1095 | TimeTraceScope timeScope("Balanced Partitioning Section Orderer" ); |
1096 | sectionOrder = runBalancedPartitioning( |
1097 | ctx, profilePath: ctx.arg.bpStartupFunctionSort ? ctx.arg.irpgoProfilePath : "" , |
1098 | forFunctionCompression: ctx.arg.bpFunctionOrderForCompression, |
1099 | forDataCompression: ctx.arg.bpDataOrderForCompression, |
1100 | compressionSortStartupFunctions: ctx.arg.bpCompressionSortStartupFunctions, |
1101 | verbose: ctx.arg.bpVerboseSectionOrderer); |
1102 | } else if (!ctx.arg.callGraphProfile.empty()) { |
1103 | sectionOrder = computeCallGraphProfileOrder(ctx); |
1104 | } |
1105 | |
1106 | if (ctx.arg.symbolOrderingFile.empty()) |
1107 | return sectionOrder; |
1108 | |
1109 | struct SymbolOrderEntry { |
1110 | int priority; |
1111 | bool present; |
1112 | }; |
1113 | |
1114 | // Build a map from symbols to their priorities. Symbols that didn't |
1115 | // appear in the symbol ordering file have the lowest priority 0. |
1116 | // All explicitly mentioned symbols have negative (higher) priorities. |
1117 | DenseMap<CachedHashStringRef, SymbolOrderEntry> symbolOrder; |
1118 | int priority = -sectionOrder.size() - ctx.arg.symbolOrderingFile.size(); |
1119 | for (StringRef s : ctx.arg.symbolOrderingFile) |
1120 | symbolOrder.insert(KV: {CachedHashStringRef(s), {.priority: priority++, .present: false}}); |
1121 | |
1122 | // Build a map from sections to their priorities. |
1123 | auto addSym = [&](Symbol &sym) { |
1124 | auto it = symbolOrder.find(Val: CachedHashStringRef(sym.getName())); |
1125 | if (it == symbolOrder.end()) |
1126 | return; |
1127 | SymbolOrderEntry &ent = it->second; |
1128 | ent.present = true; |
1129 | |
1130 | maybeWarnUnorderableSymbol(ctx, sym: &sym); |
1131 | |
1132 | if (auto *d = dyn_cast<Defined>(Val: &sym)) { |
1133 | if (auto *sec = dyn_cast_or_null<InputSectionBase>(Val: d->section)) { |
1134 | int &priority = sectionOrder[cast<InputSectionBase>(Val: sec)]; |
1135 | priority = std::min(a: priority, b: ent.priority); |
1136 | } |
1137 | } |
1138 | }; |
1139 | |
1140 | // We want both global and local symbols. We get the global ones from the |
1141 | // symbol table and iterate the object files for the local ones. |
1142 | for (Symbol *sym : ctx.symtab->getSymbols()) |
1143 | addSym(*sym); |
1144 | |
1145 | for (ELFFileBase *file : ctx.objectFiles) |
1146 | for (Symbol *sym : file->getLocalSymbols()) |
1147 | addSym(*sym); |
1148 | |
1149 | if (ctx.arg.warnSymbolOrdering) |
1150 | for (auto orderEntry : symbolOrder) |
1151 | if (!orderEntry.second.present) |
1152 | Warn(ctx) << "symbol ordering file: no such symbol: " |
1153 | << orderEntry.first.val(); |
1154 | |
1155 | return sectionOrder; |
1156 | } |
1157 | |
1158 | // Sorts the sections in ISD according to the provided section order. |
1159 | static void |
1160 | sortISDBySectionOrder(Ctx &ctx, InputSectionDescription *isd, |
1161 | const DenseMap<const InputSectionBase *, int> &order, |
1162 | bool executableOutputSection) { |
1163 | SmallVector<InputSection *, 0> unorderedSections; |
1164 | SmallVector<std::pair<InputSection *, int>, 0> orderedSections; |
1165 | uint64_t unorderedSize = 0; |
1166 | uint64_t totalSize = 0; |
1167 | |
1168 | for (InputSection *isec : isd->sections) { |
1169 | if (executableOutputSection) |
1170 | totalSize += isec->getSize(); |
1171 | auto i = order.find(Val: isec); |
1172 | if (i == order.end()) { |
1173 | unorderedSections.push_back(Elt: isec); |
1174 | unorderedSize += isec->getSize(); |
1175 | continue; |
1176 | } |
1177 | orderedSections.push_back(Elt: {isec, i->second}); |
1178 | } |
1179 | llvm::sort(C&: orderedSections, Comp: llvm::less_second()); |
1180 | |
1181 | // Find an insertion point for the ordered section list in the unordered |
1182 | // section list. On targets with limited-range branches, this is the mid-point |
1183 | // of the unordered section list. This decreases the likelihood that a range |
1184 | // extension thunk will be needed to enter or exit the ordered region. If the |
1185 | // ordered section list is a list of hot functions, we can generally expect |
1186 | // the ordered functions to be called more often than the unordered functions, |
1187 | // making it more likely that any particular call will be within range, and |
1188 | // therefore reducing the number of thunks required. |
1189 | // |
1190 | // For example, imagine that you have 8MB of hot code and 32MB of cold code. |
1191 | // If the layout is: |
1192 | // |
1193 | // 8MB hot |
1194 | // 32MB cold |
1195 | // |
1196 | // only the first 8-16MB of the cold code (depending on which hot function it |
1197 | // is actually calling) can call the hot code without a range extension thunk. |
1198 | // However, if we use this layout: |
1199 | // |
1200 | // 16MB cold |
1201 | // 8MB hot |
1202 | // 16MB cold |
1203 | // |
1204 | // both the last 8-16MB of the first block of cold code and the first 8-16MB |
1205 | // of the second block of cold code can call the hot code without a thunk. So |
1206 | // we effectively double the amount of code that could potentially call into |
1207 | // the hot code without a thunk. |
1208 | // |
1209 | // The above is not necessary if total size of input sections in this "isd" |
1210 | // is small. Note that we assume all input sections are executable if the |
1211 | // output section is executable (which is not always true but supposed to |
1212 | // cover most cases). |
1213 | size_t insPt = 0; |
1214 | if (executableOutputSection && !orderedSections.empty() && |
1215 | ctx.target->getThunkSectionSpacing() && |
1216 | totalSize >= ctx.target->getThunkSectionSpacing()) { |
1217 | uint64_t unorderedPos = 0; |
1218 | for (; insPt != unorderedSections.size(); ++insPt) { |
1219 | unorderedPos += unorderedSections[insPt]->getSize(); |
1220 | if (unorderedPos > unorderedSize / 2) |
1221 | break; |
1222 | } |
1223 | } |
1224 | |
1225 | isd->sections.clear(); |
1226 | for (InputSection *isec : ArrayRef(unorderedSections).slice(N: 0, M: insPt)) |
1227 | isd->sections.push_back(Elt: isec); |
1228 | for (std::pair<InputSection *, int> p : orderedSections) |
1229 | isd->sections.push_back(Elt: p.first); |
1230 | for (InputSection *isec : ArrayRef(unorderedSections).slice(N: insPt)) |
1231 | isd->sections.push_back(Elt: isec); |
1232 | } |
1233 | |
1234 | static void sortSection(Ctx &ctx, OutputSection &osec, |
1235 | const DenseMap<const InputSectionBase *, int> &order) { |
1236 | StringRef name = osec.name; |
1237 | |
1238 | // Never sort these. |
1239 | if (name == ".init" || name == ".fini" ) |
1240 | return; |
1241 | |
1242 | // Sort input sections by priority using the list provided by |
1243 | // --symbol-ordering-file or --shuffle-sections=. This is a least significant |
1244 | // digit radix sort. The sections may be sorted stably again by a more |
1245 | // significant key. |
1246 | if (!order.empty()) |
1247 | for (SectionCommand *b : osec.commands) |
1248 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: b)) |
1249 | sortISDBySectionOrder(ctx, isd, order, executableOutputSection: osec.flags & SHF_EXECINSTR); |
1250 | |
1251 | if (ctx.script->hasSectionsCommand) |
1252 | return; |
1253 | |
1254 | if (name == ".init_array" || name == ".fini_array" ) { |
1255 | osec.sortInitFini(); |
1256 | } else if (name == ".ctors" || name == ".dtors" ) { |
1257 | osec.sortCtorsDtors(); |
1258 | } else if (ctx.arg.emachine == EM_PPC64 && name == ".toc" ) { |
1259 | // .toc is allocated just after .got and is accessed using GOT-relative |
1260 | // relocations. Object files compiled with small code model have an |
1261 | // addressable range of [.got, .got + 0xFFFC] for GOT-relative relocations. |
1262 | // To reduce the risk of relocation overflow, .toc contents are sorted so |
1263 | // that sections having smaller relocation offsets are at beginning of .toc |
1264 | assert(osec.commands.size() == 1); |
1265 | auto *isd = cast<InputSectionDescription>(Val: osec.commands[0]); |
1266 | llvm::stable_sort(Range&: isd->sections, |
1267 | C: [](const InputSection *a, const InputSection *b) -> bool { |
1268 | return a->file->ppc64SmallCodeModelTocRelocs && |
1269 | !b->file->ppc64SmallCodeModelTocRelocs; |
1270 | }); |
1271 | } |
1272 | } |
1273 | |
1274 | // Sort sections within each InputSectionDescription. |
1275 | template <class ELFT> void Writer<ELFT>::sortInputSections() { |
1276 | // Assign negative priorities. |
1277 | DenseMap<const InputSectionBase *, int> order = buildSectionOrder(ctx); |
1278 | // Assign non-negative priorities due to --shuffle-sections. |
1279 | maybeShuffle(ctx, order); |
1280 | for (SectionCommand *cmd : ctx.script->sectionCommands) |
1281 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) |
1282 | sortSection(ctx, osec&: osd->osec, order); |
1283 | } |
1284 | |
1285 | template <class ELFT> void Writer<ELFT>::sortSections() { |
1286 | llvm::TimeTraceScope timeScope("Sort sections" ); |
1287 | |
1288 | // Don't sort if using -r. It is not necessary and we want to preserve the |
1289 | // relative order for SHF_LINK_ORDER sections. |
1290 | if (ctx.arg.relocatable) { |
1291 | ctx.script->adjustOutputSections(); |
1292 | return; |
1293 | } |
1294 | |
1295 | sortInputSections(); |
1296 | |
1297 | for (SectionCommand *cmd : ctx.script->sectionCommands) |
1298 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) |
1299 | osd->osec.sortRank = getSectionRank(ctx, osec&: osd->osec); |
1300 | if (!ctx.script->hasSectionsCommand) { |
1301 | // OutputDescs are mostly contiguous, but may be interleaved with |
1302 | // SymbolAssignments in the presence of INSERT commands. |
1303 | auto mid = std::stable_partition( |
1304 | ctx.script->sectionCommands.begin(), ctx.script->sectionCommands.end(), |
1305 | [](SectionCommand *cmd) { return isa<OutputDesc>(Val: cmd); }); |
1306 | std::stable_sort( |
1307 | ctx.script->sectionCommands.begin(), mid, |
1308 | [&ctx = ctx](auto *l, auto *r) { return compareSections(ctx, l, r); }); |
1309 | } |
1310 | |
1311 | // Process INSERT commands and update output section attributes. From this |
1312 | // point onwards the order of script->sectionCommands is fixed. |
1313 | ctx.script->processInsertCommands(); |
1314 | ctx.script->adjustOutputSections(); |
1315 | |
1316 | if (ctx.script->hasSectionsCommand) |
1317 | sortOrphanSections(); |
1318 | |
1319 | ctx.script->adjustSectionsAfterSorting(); |
1320 | } |
1321 | |
1322 | template <class ELFT> void Writer<ELFT>::sortOrphanSections() { |
1323 | // Orphan sections are sections present in the input files which are |
1324 | // not explicitly placed into the output file by the linker script. |
1325 | // |
1326 | // The sections in the linker script are already in the correct |
1327 | // order. We have to figuere out where to insert the orphan |
1328 | // sections. |
1329 | // |
1330 | // The order of the sections in the script is arbitrary and may not agree with |
1331 | // compareSections. This means that we cannot easily define a strict weak |
1332 | // ordering. To see why, consider a comparison of a section in the script and |
1333 | // one not in the script. We have a two simple options: |
1334 | // * Make them equivalent (a is not less than b, and b is not less than a). |
1335 | // The problem is then that equivalence has to be transitive and we can |
1336 | // have sections a, b and c with only b in a script and a less than c |
1337 | // which breaks this property. |
1338 | // * Use compareSectionsNonScript. Given that the script order doesn't have |
1339 | // to match, we can end up with sections a, b, c, d where b and c are in the |
1340 | // script and c is compareSectionsNonScript less than b. In which case d |
1341 | // can be equivalent to c, a to b and d < a. As a concrete example: |
1342 | // .a (rx) # not in script |
1343 | // .b (rx) # in script |
1344 | // .c (ro) # in script |
1345 | // .d (ro) # not in script |
1346 | // |
1347 | // The way we define an order then is: |
1348 | // * Sort only the orphan sections. They are in the end right now. |
1349 | // * Move each orphan section to its preferred position. We try |
1350 | // to put each section in the last position where it can share |
1351 | // a PT_LOAD. |
1352 | // |
1353 | // There is some ambiguity as to where exactly a new entry should be |
1354 | // inserted, because Commands contains not only output section |
1355 | // commands but also other types of commands such as symbol assignment |
1356 | // expressions. There's no correct answer here due to the lack of the |
1357 | // formal specification of the linker script. We use heuristics to |
1358 | // determine whether a new output command should be added before or |
1359 | // after another commands. For the details, look at shouldSkip |
1360 | // function. |
1361 | |
1362 | auto i = ctx.script->sectionCommands.begin(); |
1363 | auto e = ctx.script->sectionCommands.end(); |
1364 | auto nonScriptI = std::find_if(i, e, [](SectionCommand *cmd) { |
1365 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) |
1366 | return osd->osec.sectionIndex == UINT32_MAX; |
1367 | return false; |
1368 | }); |
1369 | |
1370 | // Sort the orphan sections. |
1371 | std::stable_sort(nonScriptI, e, [&ctx = ctx](auto *l, auto *r) { |
1372 | return compareSections(ctx, l, r); |
1373 | }); |
1374 | |
1375 | // As a horrible special case, skip the first . assignment if it is before any |
1376 | // section. We do this because it is common to set a load address by starting |
1377 | // the script with ". = 0xabcd" and the expectation is that every section is |
1378 | // after that. |
1379 | auto firstSectionOrDotAssignment = |
1380 | std::find_if(i, e, [](SectionCommand *cmd) { return !shouldSkip(cmd); }); |
1381 | if (firstSectionOrDotAssignment != e && |
1382 | isa<SymbolAssignment>(**firstSectionOrDotAssignment)) |
1383 | ++firstSectionOrDotAssignment; |
1384 | i = firstSectionOrDotAssignment; |
1385 | |
1386 | while (nonScriptI != e) { |
1387 | auto pos = findOrphanPos(ctx, i, nonScriptI); |
1388 | OutputSection *orphan = &cast<OutputDesc>(*nonScriptI)->osec; |
1389 | |
1390 | // As an optimization, find all sections with the same sort rank |
1391 | // and insert them with one rotate. |
1392 | unsigned rank = orphan->sortRank; |
1393 | auto end = std::find_if(nonScriptI + 1, e, [=](SectionCommand *cmd) { |
1394 | return cast<OutputDesc>(Val: cmd)->osec.sortRank != rank; |
1395 | }); |
1396 | std::rotate(pos, nonScriptI, end); |
1397 | nonScriptI = end; |
1398 | } |
1399 | } |
1400 | |
1401 | static bool compareByFilePosition(InputSection *a, InputSection *b) { |
1402 | InputSection *la = a->flags & SHF_LINK_ORDER ? a->getLinkOrderDep() : nullptr; |
1403 | InputSection *lb = b->flags & SHF_LINK_ORDER ? b->getLinkOrderDep() : nullptr; |
1404 | // SHF_LINK_ORDER sections with non-zero sh_link are ordered before |
1405 | // non-SHF_LINK_ORDER sections and SHF_LINK_ORDER sections with zero sh_link. |
1406 | if (!la || !lb) |
1407 | return la && !lb; |
1408 | OutputSection *aOut = la->getParent(); |
1409 | OutputSection *bOut = lb->getParent(); |
1410 | |
1411 | if (aOut == bOut) |
1412 | return la->outSecOff < lb->outSecOff; |
1413 | if (aOut->addr == bOut->addr) |
1414 | return aOut->sectionIndex < bOut->sectionIndex; |
1415 | return aOut->addr < bOut->addr; |
1416 | } |
1417 | |
1418 | template <class ELFT> void Writer<ELFT>::resolveShfLinkOrder() { |
1419 | llvm::TimeTraceScope timeScope("Resolve SHF_LINK_ORDER" ); |
1420 | for (OutputSection *sec : ctx.outputSections) { |
1421 | if (!(sec->flags & SHF_LINK_ORDER)) |
1422 | continue; |
1423 | |
1424 | // The ARM.exidx section use SHF_LINK_ORDER, but we have consolidated |
1425 | // this processing inside the ARMExidxsyntheticsection::finalizeContents(). |
1426 | if (!ctx.arg.relocatable && ctx.arg.emachine == EM_ARM && |
1427 | sec->type == SHT_ARM_EXIDX) |
1428 | continue; |
1429 | |
1430 | // Link order may be distributed across several InputSectionDescriptions. |
1431 | // Sorting is performed separately. |
1432 | SmallVector<InputSection **, 0> scriptSections; |
1433 | SmallVector<InputSection *, 0> sections; |
1434 | for (SectionCommand *cmd : sec->commands) { |
1435 | auto *isd = dyn_cast<InputSectionDescription>(Val: cmd); |
1436 | if (!isd) |
1437 | continue; |
1438 | bool hasLinkOrder = false; |
1439 | scriptSections.clear(); |
1440 | sections.clear(); |
1441 | for (InputSection *&isec : isd->sections) { |
1442 | if (isec->flags & SHF_LINK_ORDER) { |
1443 | InputSection *link = isec->getLinkOrderDep(); |
1444 | if (link && !link->getParent()) |
1445 | ErrAlways(ctx) << isec << ": sh_link points to discarded section " |
1446 | << link; |
1447 | hasLinkOrder = true; |
1448 | } |
1449 | scriptSections.push_back(Elt: &isec); |
1450 | sections.push_back(Elt: isec); |
1451 | } |
1452 | if (hasLinkOrder && errCount(ctx) == 0) { |
1453 | llvm::stable_sort(Range&: sections, C: compareByFilePosition); |
1454 | for (int i = 0, n = sections.size(); i != n; ++i) |
1455 | *scriptSections[i] = sections[i]; |
1456 | } |
1457 | } |
1458 | } |
1459 | } |
1460 | |
1461 | static void finalizeSynthetic(Ctx &ctx, SyntheticSection *sec) { |
1462 | if (sec && sec->isNeeded() && sec->getParent()) { |
1463 | llvm::TimeTraceScope timeScope("Finalize synthetic sections" , sec->name); |
1464 | sec->finalizeContents(); |
1465 | } |
1466 | } |
1467 | |
1468 | static bool canInsertPadding(OutputSection *sec) { |
1469 | StringRef s = sec->name; |
1470 | return s == ".bss" || s == ".data" || s == ".data.rel.ro" || s == ".lbss" || |
1471 | s == ".ldata" || s == ".lrodata" || s == ".ltext" || s == ".rodata" || |
1472 | s.starts_with(Prefix: ".text" ); |
1473 | } |
1474 | |
1475 | static void randomizeSectionPadding(Ctx &ctx) { |
1476 | std::mt19937 g(*ctx.arg.randomizeSectionPadding); |
1477 | PhdrEntry *curPtLoad = nullptr; |
1478 | for (OutputSection *os : ctx.outputSections) { |
1479 | if (!canInsertPadding(sec: os)) |
1480 | continue; |
1481 | for (SectionCommand *bc : os->commands) { |
1482 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: bc)) { |
1483 | SmallVector<InputSection *, 0> tmp; |
1484 | if (os->ptLoad != curPtLoad) { |
1485 | tmp.push_back(Elt: make<RandomizePaddingSection>( |
1486 | args&: ctx, args: g() % ctx.arg.maxPageSize, args&: os)); |
1487 | curPtLoad = os->ptLoad; |
1488 | } |
1489 | for (InputSection *isec : isd->sections) { |
1490 | // Probability of inserting padding is 1 in 16. |
1491 | if (g() % 16 == 0) |
1492 | tmp.push_back( |
1493 | Elt: make<RandomizePaddingSection>(args&: ctx, args&: isec->addralign, args&: os)); |
1494 | tmp.push_back(Elt: isec); |
1495 | } |
1496 | isd->sections = std::move(tmp); |
1497 | } |
1498 | } |
1499 | } |
1500 | } |
1501 | |
1502 | // We need to generate and finalize the content that depends on the address of |
1503 | // InputSections. As the generation of the content may also alter InputSection |
1504 | // addresses we must converge to a fixed point. We do that here. See the comment |
1505 | // in Writer<ELFT>::finalizeSections(). |
1506 | template <class ELFT> void Writer<ELFT>::finalizeAddressDependentContent() { |
1507 | llvm::TimeTraceScope timeScope("Finalize address dependent content" ); |
1508 | AArch64Err843419Patcher a64p(ctx); |
1509 | ARMErr657417Patcher a32p(ctx); |
1510 | ctx.script->assignAddresses(); |
1511 | |
1512 | // .ARM.exidx and SHF_LINK_ORDER do not require precise addresses, but they |
1513 | // do require the relative addresses of OutputSections because linker scripts |
1514 | // can assign Virtual Addresses to OutputSections that are not monotonically |
1515 | // increasing. Anything here must be repeatable, since spilling may change |
1516 | // section order. |
1517 | const auto finalizeOrderDependentContent = [this] { |
1518 | for (Partition &part : ctx.partitions) |
1519 | finalizeSynthetic(ctx, sec: part.armExidx.get()); |
1520 | resolveShfLinkOrder(); |
1521 | }; |
1522 | finalizeOrderDependentContent(); |
1523 | |
1524 | // Converts call x@GDPLT to call __tls_get_addr |
1525 | if (ctx.arg.emachine == EM_HEXAGON) |
1526 | hexagonTLSSymbolUpdate(ctx); |
1527 | |
1528 | if (ctx.arg.randomizeSectionPadding) |
1529 | randomizeSectionPadding(ctx); |
1530 | |
1531 | uint32_t pass = 0, assignPasses = 0; |
1532 | for (;;) { |
1533 | bool changed = ctx.target->needsThunks |
1534 | ? tc.createThunks(pass, outputSections: ctx.outputSections) |
1535 | : ctx.target->relaxOnce(pass); |
1536 | bool spilled = ctx.script->spillSections(); |
1537 | changed |= spilled; |
1538 | ++pass; |
1539 | |
1540 | // With Thunk Size much smaller than branch range we expect to |
1541 | // converge quickly; if we get to 30 something has gone wrong. |
1542 | if (changed && pass >= 30) { |
1543 | Err(ctx) << "address assignment did not converge" ; |
1544 | break; |
1545 | } |
1546 | |
1547 | if (ctx.arg.fixCortexA53Errata843419) { |
1548 | if (changed) |
1549 | ctx.script->assignAddresses(); |
1550 | changed |= a64p.createFixes(); |
1551 | } |
1552 | if (ctx.arg.fixCortexA8) { |
1553 | if (changed) |
1554 | ctx.script->assignAddresses(); |
1555 | changed |= a32p.createFixes(); |
1556 | } |
1557 | |
1558 | finalizeSynthetic(ctx, sec: ctx.in.got.get()); |
1559 | if (ctx.in.mipsGot) |
1560 | ctx.in.mipsGot->updateAllocSize(ctx); |
1561 | |
1562 | for (Partition &part : ctx.partitions) { |
1563 | // The R_AARCH64_AUTH_RELATIVE has a smaller addend field as bits [63:32] |
1564 | // encode the signing schema. We've put relocations in .relr.auth.dyn |
1565 | // during RelocationScanner::processAux, but the target VA for some of |
1566 | // them might be wider than 32 bits. We can only know the final VA at this |
1567 | // point, so move relocations with large values from .relr.auth.dyn to |
1568 | // .rela.dyn. See also AArch64::relocate. |
1569 | if (part.relrAuthDyn) { |
1570 | auto it = llvm::remove_if( |
1571 | part.relrAuthDyn->relocs, [this, &part](const RelativeReloc &elem) { |
1572 | const Relocation &reloc = elem.inputSec->relocs()[elem.relocIdx]; |
1573 | if (isInt<32>(x: reloc.sym->getVA(ctx, addend: reloc.addend))) |
1574 | return false; |
1575 | part.relaDyn->addReloc(reloc: {R_AARCH64_AUTH_RELATIVE, elem.inputSec, |
1576 | reloc.offset, |
1577 | DynamicReloc::AddendOnlyWithTargetVA, |
1578 | *reloc.sym, reloc.addend, R_ABS}); |
1579 | return true; |
1580 | }); |
1581 | changed |= (it != part.relrAuthDyn->relocs.end()); |
1582 | part.relrAuthDyn->relocs.erase(it, part.relrAuthDyn->relocs.end()); |
1583 | } |
1584 | if (part.relaDyn) |
1585 | changed |= part.relaDyn->updateAllocSize(ctx); |
1586 | if (part.relrDyn) |
1587 | changed |= part.relrDyn->updateAllocSize(ctx); |
1588 | if (part.relrAuthDyn) |
1589 | changed |= part.relrAuthDyn->updateAllocSize(ctx); |
1590 | if (part.memtagGlobalDescriptors) |
1591 | changed |= part.memtagGlobalDescriptors->updateAllocSize(ctx); |
1592 | } |
1593 | |
1594 | std::pair<const OutputSection *, const Defined *> changes = |
1595 | ctx.script->assignAddresses(); |
1596 | if (!changed) { |
1597 | // Some symbols may be dependent on section addresses. When we break the |
1598 | // loop, the symbol values are finalized because a previous |
1599 | // assignAddresses() finalized section addresses. |
1600 | if (!changes.first && !changes.second) |
1601 | break; |
1602 | if (++assignPasses == 5) { |
1603 | if (changes.first) |
1604 | Err(ctx) << "address (0x" << Twine::utohexstr(Val: changes.first->addr) |
1605 | << ") of section '" << changes.first->name |
1606 | << "' does not converge" ; |
1607 | if (changes.second) |
1608 | Err(ctx) << "assignment to symbol " << changes.second |
1609 | << " does not converge" ; |
1610 | break; |
1611 | } |
1612 | } else if (spilled) { |
1613 | // Spilling can change relative section order. |
1614 | finalizeOrderDependentContent(); |
1615 | } |
1616 | } |
1617 | if (!ctx.arg.relocatable) |
1618 | ctx.target->finalizeRelax(passes: pass); |
1619 | |
1620 | if (ctx.arg.relocatable) |
1621 | for (OutputSection *sec : ctx.outputSections) |
1622 | sec->addr = 0; |
1623 | |
1624 | uint64_t imageBase = ctx.script->hasSectionsCommand || ctx.arg.relocatable |
1625 | ? 0 |
1626 | : ctx.target->getImageBase(); |
1627 | for (SectionCommand *cmd : ctx.script->sectionCommands) { |
1628 | auto *osd = dyn_cast<OutputDesc>(Val: cmd); |
1629 | if (!osd) |
1630 | continue; |
1631 | OutputSection *osec = &osd->osec; |
1632 | // Error if the address is below the image base when SECTIONS is absent |
1633 | // (e.g. when -Ttext is specified and smaller than the default target image |
1634 | // base for no-pie). |
1635 | if (osec->addr < imageBase && (osec->flags & SHF_ALLOC)) { |
1636 | Err(ctx) << "section '" << osec->name << "' address (0x" |
1637 | << Twine::utohexstr(Val: osec->addr) |
1638 | << ") is smaller than image base (0x" |
1639 | << Twine::utohexstr(Val: imageBase) << "); specify --image-base" ; |
1640 | } |
1641 | |
1642 | // If addrExpr is set, the address may not be a multiple of the alignment. |
1643 | // Warn because this is error-prone. |
1644 | if (osec->addr % osec->addralign != 0) |
1645 | Warn(ctx) << "address (0x" << Twine::utohexstr(Val: osec->addr) |
1646 | << ") of section " << osec->name |
1647 | << " is not a multiple of alignment (" << osec->addralign |
1648 | << ")" ; |
1649 | } |
1650 | |
1651 | // Sizes are no longer allowed to grow, so all allowable spills have been |
1652 | // taken. Remove any leftover potential spills. |
1653 | ctx.script->erasePotentialSpillSections(); |
1654 | } |
1655 | |
1656 | // If Input Sections have been shrunk (basic block sections) then |
1657 | // update symbol values and sizes associated with these sections. With basic |
1658 | // block sections, input sections can shrink when the jump instructions at |
1659 | // the end of the section are relaxed. |
1660 | static void fixSymbolsAfterShrinking(Ctx &ctx) { |
1661 | for (InputFile *File : ctx.objectFiles) { |
1662 | parallelForEach(R: File->getSymbols(), Fn: [&](Symbol *Sym) { |
1663 | auto *def = dyn_cast<Defined>(Val: Sym); |
1664 | if (!def) |
1665 | return; |
1666 | |
1667 | const SectionBase *sec = def->section; |
1668 | if (!sec) |
1669 | return; |
1670 | |
1671 | const InputSectionBase *inputSec = dyn_cast<InputSectionBase>(Val: sec); |
1672 | if (!inputSec || !inputSec->bytesDropped) |
1673 | return; |
1674 | |
1675 | const size_t OldSize = inputSec->content().size(); |
1676 | const size_t NewSize = OldSize - inputSec->bytesDropped; |
1677 | |
1678 | if (def->value > NewSize && def->value <= OldSize) { |
1679 | LLVM_DEBUG(llvm::dbgs() |
1680 | << "Moving symbol " << Sym->getName() << " from " |
1681 | << def->value << " to " |
1682 | << def->value - inputSec->bytesDropped << " bytes\n" ); |
1683 | def->value -= inputSec->bytesDropped; |
1684 | return; |
1685 | } |
1686 | |
1687 | if (def->value + def->size > NewSize && def->value <= OldSize && |
1688 | def->value + def->size <= OldSize) { |
1689 | LLVM_DEBUG(llvm::dbgs() |
1690 | << "Shrinking symbol " << Sym->getName() << " from " |
1691 | << def->size << " to " << def->size - inputSec->bytesDropped |
1692 | << " bytes\n" ); |
1693 | def->size -= inputSec->bytesDropped; |
1694 | } |
1695 | }); |
1696 | } |
1697 | } |
1698 | |
1699 | // If basic block sections exist, there are opportunities to delete fall thru |
1700 | // jumps and shrink jump instructions after basic block reordering. This |
1701 | // relaxation pass does that. It is only enabled when --optimize-bb-jumps |
1702 | // option is used. |
1703 | template <class ELFT> void Writer<ELFT>::optimizeBasicBlockJumps() { |
1704 | assert(ctx.arg.optimizeBBJumps); |
1705 | SmallVector<InputSection *, 0> storage; |
1706 | |
1707 | ctx.script->assignAddresses(); |
1708 | // For every output section that has executable input sections, this |
1709 | // does the following: |
1710 | // 1. Deletes all direct jump instructions in input sections that |
1711 | // jump to the following section as it is not required. |
1712 | // 2. If there are two consecutive jump instructions, it checks |
1713 | // if they can be flipped and one can be deleted. |
1714 | for (OutputSection *osec : ctx.outputSections) { |
1715 | if (!(osec->flags & SHF_EXECINSTR)) |
1716 | continue; |
1717 | ArrayRef<InputSection *> sections = getInputSections(os: *osec, storage); |
1718 | size_t numDeleted = 0; |
1719 | // Delete all fall through jump instructions. Also, check if two |
1720 | // consecutive jump instructions can be flipped so that a fall |
1721 | // through jmp instruction can be deleted. |
1722 | for (size_t i = 0, e = sections.size(); i != e; ++i) { |
1723 | InputSection *next = i + 1 < sections.size() ? sections[i + 1] : nullptr; |
1724 | InputSection &sec = *sections[i]; |
1725 | numDeleted += ctx.target->deleteFallThruJmpInsn(is&: sec, file: sec.file, nextIS: next); |
1726 | } |
1727 | if (numDeleted > 0) { |
1728 | ctx.script->assignAddresses(); |
1729 | LLVM_DEBUG(llvm::dbgs() |
1730 | << "Removing " << numDeleted << " fall through jumps\n" ); |
1731 | } |
1732 | } |
1733 | |
1734 | fixSymbolsAfterShrinking(ctx); |
1735 | |
1736 | for (OutputSection *osec : ctx.outputSections) |
1737 | for (InputSection *is : getInputSections(os: *osec, storage)) |
1738 | is->trim(); |
1739 | } |
1740 | |
1741 | // In order to allow users to manipulate linker-synthesized sections, |
1742 | // we had to add synthetic sections to the input section list early, |
1743 | // even before we make decisions whether they are needed. This allows |
1744 | // users to write scripts like this: ".mygot : { .got }". |
1745 | // |
1746 | // Doing it has an unintended side effects. If it turns out that we |
1747 | // don't need a .got (for example) at all because there's no |
1748 | // relocation that needs a .got, we don't want to emit .got. |
1749 | // |
1750 | // To deal with the above problem, this function is called after |
1751 | // scanRelocations is called to remove synthetic sections that turn |
1752 | // out to be empty. |
1753 | static void removeUnusedSyntheticSections(Ctx &ctx) { |
1754 | // All input synthetic sections that can be empty are placed after |
1755 | // all regular ones. Reverse iterate to find the first synthetic section |
1756 | // after a non-synthetic one which will be our starting point. |
1757 | auto start = |
1758 | llvm::find_if(Range: llvm::reverse(C&: ctx.inputSections), P: [](InputSectionBase *s) { |
1759 | return !isa<SyntheticSection>(Val: s); |
1760 | }).base(); |
1761 | |
1762 | // Remove unused synthetic sections from ctx.inputSections; |
1763 | DenseSet<InputSectionBase *> unused; |
1764 | auto end = |
1765 | std::remove_if(first: start, last: ctx.inputSections.end(), pred: [&](InputSectionBase *s) { |
1766 | auto *sec = cast<SyntheticSection>(Val: s); |
1767 | if (sec->getParent() && sec->isNeeded()) |
1768 | return false; |
1769 | // .relr.auth.dyn relocations may be moved to .rela.dyn in |
1770 | // finalizeAddressDependentContent, making .rela.dyn no longer empty. |
1771 | // Conservatively keep .rela.dyn. .relr.auth.dyn can be made empty, but |
1772 | // we would fail to remove it here. |
1773 | if (ctx.arg.emachine == EM_AARCH64 && ctx.arg.relrPackDynRelocs && |
1774 | sec == ctx.mainPart->relaDyn.get()) |
1775 | return false; |
1776 | unused.insert(V: sec); |
1777 | return true; |
1778 | }); |
1779 | ctx.inputSections.erase(CS: end, CE: ctx.inputSections.end()); |
1780 | |
1781 | // Remove unused synthetic sections from the corresponding input section |
1782 | // description and orphanSections. |
1783 | for (auto *sec : unused) |
1784 | if (OutputSection *osec = cast<SyntheticSection>(Val: sec)->getParent()) |
1785 | for (SectionCommand *cmd : osec->commands) |
1786 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: cmd)) |
1787 | llvm::erase_if(C&: isd->sections, P: [&](InputSection *isec) { |
1788 | return unused.count(V: isec); |
1789 | }); |
1790 | llvm::erase_if(C&: ctx.script->orphanSections, P: [&](const InputSectionBase *sec) { |
1791 | return unused.count(V: sec); |
1792 | }); |
1793 | } |
1794 | |
1795 | // Create output section objects and add them to OutputSections. |
1796 | template <class ELFT> void Writer<ELFT>::finalizeSections() { |
1797 | if (!ctx.arg.relocatable) { |
1798 | ctx.out.preinitArray = findSection(ctx, name: ".preinit_array" ); |
1799 | ctx.out.initArray = findSection(ctx, name: ".init_array" ); |
1800 | ctx.out.finiArray = findSection(ctx, name: ".fini_array" ); |
1801 | |
1802 | // The linker needs to define SECNAME_start, SECNAME_end and SECNAME_stop |
1803 | // symbols for sections, so that the runtime can get the start and end |
1804 | // addresses of each section by section name. Add such symbols. |
1805 | addStartEndSymbols(); |
1806 | for (SectionCommand *cmd : ctx.script->sectionCommands) |
1807 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) |
1808 | addStartStopSymbols(osec&: osd->osec); |
1809 | |
1810 | // Add _DYNAMIC symbol. Unlike GNU gold, our _DYNAMIC symbol has no type. |
1811 | // It should be okay as no one seems to care about the type. |
1812 | // Even the author of gold doesn't remember why gold behaves that way. |
1813 | // https://sourceware.org/ml/binutils/2002-03/msg00360.html |
1814 | if (ctx.mainPart->dynamic->parent) { |
1815 | Symbol *s = ctx.symtab->addSymbol(newSym: Defined{ |
1816 | ctx, ctx.internalFile, "_DYNAMIC" , STB_WEAK, STV_HIDDEN, STT_NOTYPE, |
1817 | /*value=*/0, /*size=*/0, ctx.mainPart->dynamic.get()}); |
1818 | s->isUsedInRegularObj = true; |
1819 | } |
1820 | |
1821 | // Define __rel[a]_iplt_{start,end} symbols if needed. |
1822 | addRelIpltSymbols(); |
1823 | |
1824 | // RISC-V's gp can address +/- 2 KiB, set it to .sdata + 0x800. This symbol |
1825 | // should only be defined in an executable. If .sdata does not exist, its |
1826 | // value/section does not matter but it has to be relative, so set its |
1827 | // st_shndx arbitrarily to 1 (ctx.out.elfHeader). |
1828 | if (ctx.arg.emachine == EM_RISCV) { |
1829 | if (!ctx.arg.shared) { |
1830 | OutputSection *sec = findSection(ctx, name: ".sdata" ); |
1831 | addOptionalRegular(ctx, name: "__global_pointer$" , |
1832 | sec: sec ? sec : ctx.out.elfHeader.get(), val: 0x800, |
1833 | stOther: STV_DEFAULT); |
1834 | // Set riscvGlobalPointer to be used by the optional global pointer |
1835 | // relaxation. |
1836 | if (ctx.arg.relaxGP) { |
1837 | Symbol *s = ctx.symtab->find(name: "__global_pointer$" ); |
1838 | if (s && s->isDefined()) |
1839 | ctx.sym.riscvGlobalPointer = cast<Defined>(Val: s); |
1840 | } |
1841 | } |
1842 | } |
1843 | |
1844 | if (ctx.arg.emachine == EM_386 || ctx.arg.emachine == EM_X86_64) { |
1845 | // On targets that support TLSDESC, _TLS_MODULE_BASE_ is defined in such a |
1846 | // way that: |
1847 | // |
1848 | // 1) Without relaxation: it produces a dynamic TLSDESC relocation that |
1849 | // computes 0. |
1850 | // 2) With LD->LE relaxation: _TLS_MODULE_BASE_@tpoff = 0 (lowest address |
1851 | // in the TLS block). |
1852 | // |
1853 | // 2) is special cased in @tpoff computation. To satisfy 1), we define it |
1854 | // as an absolute symbol of zero. This is different from GNU linkers which |
1855 | // define _TLS_MODULE_BASE_ relative to the first TLS section. |
1856 | Symbol *s = ctx.symtab->find(name: "_TLS_MODULE_BASE_" ); |
1857 | if (s && s->isUndefined()) { |
1858 | s->resolve(ctx, other: Defined{ctx, ctx.internalFile, StringRef(), STB_GLOBAL, |
1859 | STV_HIDDEN, STT_TLS, /*value=*/0, 0, |
1860 | /*section=*/nullptr}); |
1861 | ctx.sym.tlsModuleBase = cast<Defined>(Val: s); |
1862 | } |
1863 | } |
1864 | |
1865 | // This responsible for splitting up .eh_frame section into |
1866 | // pieces. The relocation scan uses those pieces, so this has to be |
1867 | // earlier. |
1868 | { |
1869 | llvm::TimeTraceScope timeScope("Finalize .eh_frame" ); |
1870 | for (Partition &part : ctx.partitions) |
1871 | finalizeSynthetic(ctx, sec: part.ehFrame.get()); |
1872 | } |
1873 | } |
1874 | |
1875 | // If the previous code block defines any non-hidden symbols (e.g. |
1876 | // __global_pointer$), they may be exported. |
1877 | if (ctx.arg.exportDynamic) |
1878 | for (Symbol *sym : ctx.synthesizedSymbols) |
1879 | if (sym->computeBinding(ctx) != STB_LOCAL) |
1880 | sym->isExported = true; |
1881 | |
1882 | demoteSymbolsAndComputeIsPreemptible(ctx); |
1883 | |
1884 | if (ctx.arg.copyRelocs && ctx.arg.discard != DiscardPolicy::None) |
1885 | markUsedLocalSymbols<ELFT>(ctx); |
1886 | demoteAndCopyLocalSymbols(ctx); |
1887 | |
1888 | if (ctx.arg.copyRelocs) |
1889 | addSectionSymbols(); |
1890 | |
1891 | // Change values of linker-script-defined symbols from placeholders (assigned |
1892 | // by declareSymbols) to actual definitions. |
1893 | ctx.script->processSymbolAssignments(); |
1894 | |
1895 | if (!ctx.arg.relocatable) { |
1896 | llvm::TimeTraceScope timeScope("Scan relocations" ); |
1897 | // Scan relocations. This must be done after every symbol is declared so |
1898 | // that we can correctly decide if a dynamic relocation is needed. This is |
1899 | // called after processSymbolAssignments() because it needs to know whether |
1900 | // a linker-script-defined symbol is absolute. |
1901 | scanRelocations<ELFT>(ctx); |
1902 | reportUndefinedSymbols(ctx); |
1903 | postScanRelocations(ctx); |
1904 | |
1905 | if (ctx.in.plt && ctx.in.plt->isNeeded()) |
1906 | ctx.in.plt->addSymbols(); |
1907 | if (ctx.in.iplt && ctx.in.iplt->isNeeded()) |
1908 | ctx.in.iplt->addSymbols(); |
1909 | |
1910 | if (ctx.arg.unresolvedSymbolsInShlib != UnresolvedPolicy::Ignore) { |
1911 | auto diag = |
1912 | ctx.arg.unresolvedSymbolsInShlib == UnresolvedPolicy::ReportError && |
1913 | !ctx.arg.noinhibitExec |
1914 | ? DiagLevel::Err |
1915 | : DiagLevel::Warn; |
1916 | // Error on undefined symbols in a shared object, if all of its DT_NEEDED |
1917 | // entries are seen. These cases would otherwise lead to runtime errors |
1918 | // reported by the dynamic linker. |
1919 | // |
1920 | // ld.bfd traces all DT_NEEDED to emulate the logic of the dynamic linker |
1921 | // to catch more cases. That is too much for us. Our approach resembles |
1922 | // the one used in ld.gold, achieves a good balance to be useful but not |
1923 | // too smart. |
1924 | // |
1925 | // If a DSO reference is resolved by a SharedSymbol, but the SharedSymbol |
1926 | // is overridden by a hidden visibility Defined (which is later discarded |
1927 | // due to GC), don't report the diagnostic. However, this may indicate an |
1928 | // unintended SharedSymbol. |
1929 | for (SharedFile *file : ctx.sharedFiles) { |
1930 | bool allNeededIsKnown = |
1931 | llvm::all_of(file->dtNeeded, [&](StringRef needed) { |
1932 | return ctx.symtab->soNames.count(Val: CachedHashStringRef(needed)); |
1933 | }); |
1934 | if (!allNeededIsKnown) |
1935 | continue; |
1936 | for (Symbol *sym : file->requiredSymbols) { |
1937 | if (sym->dsoDefined) |
1938 | continue; |
1939 | if (sym->isUndefined() && !sym->isWeak()) { |
1940 | ELFSyncStream(ctx, diag) |
1941 | << "undefined reference: " << sym << "\n>>> referenced by " |
1942 | << file << " (disallowed by --no-allow-shlib-undefined)" ; |
1943 | } else if (sym->isDefined() && |
1944 | sym->computeBinding(ctx) == STB_LOCAL) { |
1945 | ELFSyncStream(ctx, diag) |
1946 | << "non-exported symbol '" << sym << "' in '" << sym->file |
1947 | << "' is referenced by DSO '" << file << "'" ; |
1948 | } |
1949 | } |
1950 | } |
1951 | } |
1952 | } |
1953 | |
1954 | { |
1955 | llvm::TimeTraceScope timeScope("Add symbols to symtabs" ); |
1956 | // Now that we have defined all possible global symbols including linker- |
1957 | // synthesized ones. Visit all symbols to give the finishing touches. |
1958 | for (Symbol *sym : ctx.symtab->getSymbols()) { |
1959 | if (!sym->isUsedInRegularObj || !includeInSymtab(ctx, b: *sym)) |
1960 | continue; |
1961 | if (!ctx.arg.relocatable) |
1962 | sym->binding = sym->computeBinding(ctx); |
1963 | if (ctx.in.symTab) |
1964 | ctx.in.symTab->addSymbol(sym); |
1965 | |
1966 | // computeBinding might localize a symbol that was considered exported |
1967 | // but then synthesized as hidden (e.g. _DYNAMIC). |
1968 | if ((sym->isExported || sym->isPreemptible) && !sym->isLocal()) { |
1969 | ctx.partitions[sym->partition - 1].dynSymTab->addSymbol(sym); |
1970 | if (auto *file = dyn_cast<SharedFile>(Val: sym->file)) |
1971 | if (file->isNeeded && !sym->isUndefined()) |
1972 | addVerneed(ctx, ss&: *sym); |
1973 | } |
1974 | } |
1975 | |
1976 | // We also need to scan the dynamic relocation tables of the other |
1977 | // partitions and add any referenced symbols to the partition's dynsym. |
1978 | for (Partition &part : |
1979 | MutableArrayRef<Partition>(ctx.partitions).slice(N: 1)) { |
1980 | DenseSet<Symbol *> syms; |
1981 | for (const SymbolTableEntry &e : part.dynSymTab->getSymbols()) |
1982 | syms.insert(V: e.sym); |
1983 | for (DynamicReloc &reloc : part.relaDyn->relocs) |
1984 | if (reloc.sym && reloc.needsDynSymIndex() && |
1985 | syms.insert(V: reloc.sym).second) |
1986 | part.dynSymTab->addSymbol(sym: reloc.sym); |
1987 | } |
1988 | } |
1989 | |
1990 | if (ctx.in.mipsGot) |
1991 | ctx.in.mipsGot->build(); |
1992 | |
1993 | removeUnusedSyntheticSections(ctx); |
1994 | ctx.script->diagnoseOrphanHandling(); |
1995 | ctx.script->diagnoseMissingSGSectionAddress(); |
1996 | |
1997 | sortSections(); |
1998 | |
1999 | // Create a list of OutputSections, assign sectionIndex, and populate |
2000 | // ctx.in.shStrTab. If -z nosectionheader is specified, drop non-ALLOC |
2001 | // sections. |
2002 | for (SectionCommand *cmd : ctx.script->sectionCommands) |
2003 | if (auto *osd = dyn_cast<OutputDesc>(Val: cmd)) { |
2004 | OutputSection *osec = &osd->osec; |
2005 | if (!ctx.in.shStrTab && !(osec->flags & SHF_ALLOC)) |
2006 | continue; |
2007 | ctx.outputSections.push_back(Elt: osec); |
2008 | osec->sectionIndex = ctx.outputSections.size(); |
2009 | if (ctx.in.shStrTab) |
2010 | osec->shName = ctx.in.shStrTab->addString(s: osec->name); |
2011 | } |
2012 | |
2013 | // Prefer command line supplied address over other constraints. |
2014 | for (OutputSection *sec : ctx.outputSections) { |
2015 | auto i = ctx.arg.sectionStartMap.find(Key: sec->name); |
2016 | if (i != ctx.arg.sectionStartMap.end()) |
2017 | sec->addrExpr = [=] { return i->second; }; |
2018 | } |
2019 | |
2020 | // With the ctx.outputSections available check for GDPLT relocations |
2021 | // and add __tls_get_addr symbol if needed. |
2022 | if (ctx.arg.emachine == EM_HEXAGON && |
2023 | hexagonNeedsTLSSymbol(outputSections: ctx.outputSections)) { |
2024 | Symbol *sym = |
2025 | ctx.symtab->addSymbol(newSym: Undefined{ctx.internalFile, "__tls_get_addr" , |
2026 | STB_GLOBAL, STV_DEFAULT, STT_NOTYPE}); |
2027 | sym->isPreemptible = true; |
2028 | ctx.partitions[0].dynSymTab->addSymbol(sym); |
2029 | } |
2030 | |
2031 | // This is a bit of a hack. A value of 0 means undef, so we set it |
2032 | // to 1 to make __ehdr_start defined. The section number is not |
2033 | // particularly relevant. |
2034 | ctx.out.elfHeader->sectionIndex = 1; |
2035 | ctx.out.elfHeader->size = sizeof(typename ELFT::Ehdr); |
2036 | |
2037 | // Binary and relocatable output does not have PHDRS. |
2038 | // The headers have to be created before finalize as that can influence the |
2039 | // image base and the dynamic section on mips includes the image base. |
2040 | if (!ctx.arg.relocatable && !ctx.arg.oFormatBinary) { |
2041 | for (Partition &part : ctx.partitions) { |
2042 | part.phdrs = ctx.script->hasPhdrsCommands() ? ctx.script->createPhdrs() |
2043 | : createPhdrs(part); |
2044 | if (ctx.arg.emachine == EM_ARM) { |
2045 | // PT_ARM_EXIDX is the ARM EHABI equivalent of PT_GNU_EH_FRAME |
2046 | addPhdrForSection(part, shType: SHT_ARM_EXIDX, pType: PT_ARM_EXIDX, pFlags: PF_R); |
2047 | } |
2048 | if (ctx.arg.emachine == EM_MIPS) { |
2049 | // Add separate segments for MIPS-specific sections. |
2050 | addPhdrForSection(part, shType: SHT_MIPS_REGINFO, pType: PT_MIPS_REGINFO, pFlags: PF_R); |
2051 | addPhdrForSection(part, shType: SHT_MIPS_OPTIONS, pType: PT_MIPS_OPTIONS, pFlags: PF_R); |
2052 | addPhdrForSection(part, shType: SHT_MIPS_ABIFLAGS, pType: PT_MIPS_ABIFLAGS, pFlags: PF_R); |
2053 | } |
2054 | if (ctx.arg.emachine == EM_RISCV) |
2055 | addPhdrForSection(part, shType: SHT_RISCV_ATTRIBUTES, pType: PT_RISCV_ATTRIBUTES, |
2056 | pFlags: PF_R); |
2057 | } |
2058 | ctx.out.programHeaders->size = |
2059 | sizeof(Elf_Phdr) * ctx.mainPart->phdrs.size(); |
2060 | |
2061 | // Find the TLS segment. This happens before the section layout loop so that |
2062 | // Android relocation packing can look up TLS symbol addresses. We only need |
2063 | // to care about the main partition here because all TLS symbols were moved |
2064 | // to the main partition (see MarkLive.cpp). |
2065 | for (auto &p : ctx.mainPart->phdrs) |
2066 | if (p->p_type == PT_TLS) |
2067 | ctx.tlsPhdr = p.get(); |
2068 | } |
2069 | |
2070 | // Some symbols are defined in term of program headers. Now that we |
2071 | // have the headers, we can find out which sections they point to. |
2072 | setReservedSymbolSections(); |
2073 | |
2074 | if (ctx.script->noCrossRefs.size()) { |
2075 | llvm::TimeTraceScope timeScope("Check NOCROSSREFS" ); |
2076 | checkNoCrossRefs<ELFT>(ctx); |
2077 | } |
2078 | |
2079 | { |
2080 | llvm::TimeTraceScope timeScope("Finalize synthetic sections" ); |
2081 | |
2082 | finalizeSynthetic(ctx, sec: ctx.in.bss.get()); |
2083 | finalizeSynthetic(ctx, sec: ctx.in.bssRelRo.get()); |
2084 | finalizeSynthetic(ctx, sec: ctx.in.symTabShndx.get()); |
2085 | finalizeSynthetic(ctx, sec: ctx.in.shStrTab.get()); |
2086 | finalizeSynthetic(ctx, sec: ctx.in.strTab.get()); |
2087 | finalizeSynthetic(ctx, sec: ctx.in.got.get()); |
2088 | finalizeSynthetic(ctx, sec: ctx.in.mipsGot.get()); |
2089 | finalizeSynthetic(ctx, sec: ctx.in.igotPlt.get()); |
2090 | finalizeSynthetic(ctx, sec: ctx.in.gotPlt.get()); |
2091 | finalizeSynthetic(ctx, sec: ctx.in.relaPlt.get()); |
2092 | finalizeSynthetic(ctx, sec: ctx.in.plt.get()); |
2093 | finalizeSynthetic(ctx, sec: ctx.in.iplt.get()); |
2094 | finalizeSynthetic(ctx, sec: ctx.in.ppc32Got2.get()); |
2095 | finalizeSynthetic(ctx, sec: ctx.in.partIndex.get()); |
2096 | |
2097 | // Dynamic section must be the last one in this list and dynamic |
2098 | // symbol table section (dynSymTab) must be the first one. |
2099 | for (Partition &part : ctx.partitions) { |
2100 | if (part.relaDyn) { |
2101 | part.relaDyn->mergeRels(); |
2102 | // Compute DT_RELACOUNT to be used by part.dynamic. |
2103 | part.relaDyn->partitionRels(); |
2104 | finalizeSynthetic(ctx, sec: part.relaDyn.get()); |
2105 | } |
2106 | if (part.relrDyn) { |
2107 | part.relrDyn->mergeRels(); |
2108 | finalizeSynthetic(ctx, sec: part.relrDyn.get()); |
2109 | } |
2110 | if (part.relrAuthDyn) { |
2111 | part.relrAuthDyn->mergeRels(); |
2112 | finalizeSynthetic(ctx, sec: part.relrAuthDyn.get()); |
2113 | } |
2114 | |
2115 | finalizeSynthetic(ctx, sec: part.dynSymTab.get()); |
2116 | finalizeSynthetic(ctx, sec: part.gnuHashTab.get()); |
2117 | finalizeSynthetic(ctx, sec: part.hashTab.get()); |
2118 | finalizeSynthetic(ctx, sec: part.verDef.get()); |
2119 | finalizeSynthetic(ctx, sec: part.ehFrameHdr.get()); |
2120 | finalizeSynthetic(ctx, sec: part.verSym.get()); |
2121 | finalizeSynthetic(ctx, sec: part.verNeed.get()); |
2122 | finalizeSynthetic(ctx, sec: part.dynamic.get()); |
2123 | } |
2124 | } |
2125 | |
2126 | if (!ctx.script->hasSectionsCommand && !ctx.arg.relocatable) |
2127 | fixSectionAlignments(); |
2128 | |
2129 | // This is used to: |
2130 | // 1) Create "thunks": |
2131 | // Jump instructions in many ISAs have small displacements, and therefore |
2132 | // they cannot jump to arbitrary addresses in memory. For example, RISC-V |
2133 | // JAL instruction can target only +-1 MiB from PC. It is a linker's |
2134 | // responsibility to create and insert small pieces of code between |
2135 | // sections to extend the ranges if jump targets are out of range. Such |
2136 | // code pieces are called "thunks". |
2137 | // |
2138 | // We add thunks at this stage. We couldn't do this before this point |
2139 | // because this is the earliest point where we know sizes of sections and |
2140 | // their layouts (that are needed to determine if jump targets are in |
2141 | // range). |
2142 | // |
2143 | // 2) Update the sections. We need to generate content that depends on the |
2144 | // address of InputSections. For example, MIPS GOT section content or |
2145 | // android packed relocations sections content. |
2146 | // |
2147 | // 3) Assign the final values for the linker script symbols. Linker scripts |
2148 | // sometimes using forward symbol declarations. We want to set the correct |
2149 | // values. They also might change after adding the thunks. |
2150 | finalizeAddressDependentContent(); |
2151 | |
2152 | // All information needed for OutputSection part of Map file is available. |
2153 | if (errCount(ctx)) |
2154 | return; |
2155 | |
2156 | { |
2157 | llvm::TimeTraceScope timeScope("Finalize synthetic sections" ); |
2158 | // finalizeAddressDependentContent may have added local symbols to the |
2159 | // static symbol table. |
2160 | finalizeSynthetic(ctx, sec: ctx.in.symTab.get()); |
2161 | finalizeSynthetic(ctx, sec: ctx.in.debugNames.get()); |
2162 | finalizeSynthetic(ctx, sec: ctx.in.ppc64LongBranchTarget.get()); |
2163 | finalizeSynthetic(ctx, sec: ctx.in.armCmseSGSection.get()); |
2164 | } |
2165 | |
2166 | // Relaxation to delete inter-basic block jumps created by basic block |
2167 | // sections. Run after ctx.in.symTab is finalized as optimizeBasicBlockJumps |
2168 | // can relax jump instructions based on symbol offset. |
2169 | if (ctx.arg.optimizeBBJumps) |
2170 | optimizeBasicBlockJumps(); |
2171 | |
2172 | // Fill other section headers. The dynamic table is finalized |
2173 | // at the end because some tags like RELSZ depend on result |
2174 | // of finalizing other sections. |
2175 | for (OutputSection *sec : ctx.outputSections) |
2176 | sec->finalize(ctx); |
2177 | |
2178 | ctx.script->checkFinalScriptConditions(); |
2179 | |
2180 | if (ctx.arg.emachine == EM_ARM && !ctx.arg.isLE && ctx.arg.armBe8) { |
2181 | addArmInputSectionMappingSymbols(ctx); |
2182 | sortArmMappingSymbols(ctx); |
2183 | } |
2184 | } |
2185 | |
2186 | // Ensure data sections are not mixed with executable sections when |
2187 | // --execute-only is used. --execute-only make pages executable but not |
2188 | // readable. |
2189 | template <class ELFT> void Writer<ELFT>::checkExecuteOnly() { |
2190 | if (!ctx.arg.executeOnly) |
2191 | return; |
2192 | |
2193 | SmallVector<InputSection *, 0> storage; |
2194 | for (OutputSection *osec : ctx.outputSections) |
2195 | if (osec->flags & SHF_EXECINSTR) |
2196 | for (InputSection *isec : getInputSections(os: *osec, storage)) |
2197 | if (!(isec->flags & SHF_EXECINSTR)) |
2198 | ErrAlways(ctx) << "cannot place " << isec << " into " << osec->name |
2199 | << ": --execute-only does not support intermingling " |
2200 | "data and code" ; |
2201 | } |
2202 | |
2203 | // Check which input sections of RX output sections don't have the |
2204 | // SHF_AARCH64_PURECODE or SHF_ARM_PURECODE flag set. |
2205 | template <class ELFT> void Writer<ELFT>::checkExecuteOnlyReport() { |
2206 | if (ctx.arg.zExecuteOnlyReport == ReportPolicy::None) |
2207 | return; |
2208 | |
2209 | auto reportUnless = [&](bool cond) -> ELFSyncStream { |
2210 | if (cond) |
2211 | return {ctx, DiagLevel::None}; |
2212 | return {ctx, toDiagLevel(policy: ctx.arg.zExecuteOnlyReport)}; |
2213 | }; |
2214 | |
2215 | uint64_t purecodeFlag = |
2216 | ctx.arg.emachine == EM_AARCH64 ? SHF_AARCH64_PURECODE : SHF_ARM_PURECODE; |
2217 | StringRef purecodeFlagName = ctx.arg.emachine == EM_AARCH64 |
2218 | ? "SHF_AARCH64_PURECODE" |
2219 | : "SHF_ARM_PURECODE" ; |
2220 | SmallVector<InputSection *, 0> storage; |
2221 | for (OutputSection *osec : ctx.outputSections) { |
2222 | if (osec->getPhdrFlags() != (PF_R | PF_X)) |
2223 | continue; |
2224 | for (InputSection *sec : getInputSections(os: *osec, storage)) { |
2225 | if (isa<SyntheticSection>(Val: sec)) |
2226 | continue; |
2227 | reportUnless(sec->flags & purecodeFlag) |
2228 | << "-z execute-only-report: " << sec << " does not have " |
2229 | << purecodeFlagName << " flag set" ; |
2230 | } |
2231 | } |
2232 | } |
2233 | |
2234 | // The linker is expected to define SECNAME_start and SECNAME_end |
2235 | // symbols for a few sections. This function defines them. |
2236 | template <class ELFT> void Writer<ELFT>::addStartEndSymbols() { |
2237 | // If the associated output section does not exist, there is ambiguity as to |
2238 | // how we define _start and _end symbols for an init/fini section. Users |
2239 | // expect no "undefined symbol" linker errors and loaders expect equal |
2240 | // st_value but do not particularly care whether the symbols are defined or |
2241 | // not. We retain the output section so that the section indexes will be |
2242 | // correct. |
2243 | auto define = [=](StringRef start, StringRef end, OutputSection *os) { |
2244 | if (os) { |
2245 | Defined *startSym = addOptionalRegular(ctx, name: start, sec: os, val: 0); |
2246 | Defined *stopSym = addOptionalRegular(ctx, name: end, sec: os, val: -1); |
2247 | if (startSym || stopSym) |
2248 | os->usedInExpression = true; |
2249 | } else { |
2250 | addOptionalRegular(ctx, name: start, sec: ctx.out.elfHeader.get(), val: 0); |
2251 | addOptionalRegular(ctx, name: end, sec: ctx.out.elfHeader.get(), val: 0); |
2252 | } |
2253 | }; |
2254 | |
2255 | define("__preinit_array_start" , "__preinit_array_end" , ctx.out.preinitArray); |
2256 | define("__init_array_start" , "__init_array_end" , ctx.out.initArray); |
2257 | define("__fini_array_start" , "__fini_array_end" , ctx.out.finiArray); |
2258 | |
2259 | // As a special case, don't unnecessarily retain .ARM.exidx, which would |
2260 | // create an empty PT_ARM_EXIDX. |
2261 | if (OutputSection *sec = findSection(ctx, name: ".ARM.exidx" )) |
2262 | define("__exidx_start" , "__exidx_end" , sec); |
2263 | } |
2264 | |
2265 | // If a section name is valid as a C identifier (which is rare because of |
2266 | // the leading '.'), linkers are expected to define __start_<secname> and |
2267 | // __stop_<secname> symbols. They are at beginning and end of the section, |
2268 | // respectively. This is not requested by the ELF standard, but GNU ld and |
2269 | // gold provide the feature, and used by many programs. |
2270 | template <class ELFT> |
2271 | void Writer<ELFT>::addStartStopSymbols(OutputSection &osec) { |
2272 | StringRef s = osec.name; |
2273 | if (!isValidCIdentifier(s)) |
2274 | return; |
2275 | StringSaver &ss = ctx.saver; |
2276 | Defined *startSym = addOptionalRegular(ctx, name: ss.save(S: "__start_" + s), sec: &osec, val: 0, |
2277 | stOther: ctx.arg.zStartStopVisibility); |
2278 | Defined *stopSym = addOptionalRegular(ctx, name: ss.save(S: "__stop_" + s), sec: &osec, val: -1, |
2279 | stOther: ctx.arg.zStartStopVisibility); |
2280 | if (startSym || stopSym) |
2281 | osec.usedInExpression = true; |
2282 | } |
2283 | |
2284 | static bool needsPtLoad(OutputSection *sec) { |
2285 | if (!(sec->flags & SHF_ALLOC)) |
2286 | return false; |
2287 | |
2288 | // Don't allocate VA space for TLS NOBITS sections. The PT_TLS PHDR is |
2289 | // responsible for allocating space for them, not the PT_LOAD that |
2290 | // contains the TLS initialization image. |
2291 | if ((sec->flags & SHF_TLS) && sec->type == SHT_NOBITS) |
2292 | return false; |
2293 | return true; |
2294 | } |
2295 | |
2296 | // Adjust phdr flags according to certain options. |
2297 | static uint64_t computeFlags(Ctx &ctx, uint64_t flags) { |
2298 | if (ctx.arg.omagic) |
2299 | return PF_R | PF_W | PF_X; |
2300 | if (ctx.arg.executeOnly && (flags & PF_X)) |
2301 | return flags & ~PF_R; |
2302 | return flags; |
2303 | } |
2304 | |
2305 | // Decide which program headers to create and which sections to include in each |
2306 | // one. |
2307 | template <class ELFT> |
2308 | SmallVector<std::unique_ptr<PhdrEntry>, 0> |
2309 | Writer<ELFT>::createPhdrs(Partition &part) { |
2310 | SmallVector<std::unique_ptr<PhdrEntry>, 0> ret; |
2311 | auto addHdr = [&, &ctx = ctx](unsigned type, unsigned flags) -> PhdrEntry * { |
2312 | ret.push_back(Elt: std::make_unique<PhdrEntry>(args&: ctx, args&: type, args&: flags)); |
2313 | return ret.back().get(); |
2314 | }; |
2315 | |
2316 | unsigned partNo = part.getNumber(ctx); |
2317 | bool isMain = partNo == 1; |
2318 | |
2319 | // Add the first PT_LOAD segment for regular output sections. |
2320 | uint64_t flags = computeFlags(ctx, flags: PF_R); |
2321 | PhdrEntry *load = nullptr; |
2322 | |
2323 | // nmagic or omagic output does not have PT_PHDR, PT_INTERP, or the readonly |
2324 | // PT_LOAD. |
2325 | if (!ctx.arg.nmagic && !ctx.arg.omagic) { |
2326 | // The first phdr entry is PT_PHDR which describes the program header |
2327 | // itself. |
2328 | if (isMain) |
2329 | addHdr(PT_PHDR, PF_R)->add(ctx.out.programHeaders.get()); |
2330 | else |
2331 | addHdr(PT_PHDR, PF_R)->add(part.programHeaders->getParent()); |
2332 | |
2333 | // PT_INTERP must be the second entry if exists. |
2334 | if (OutputSection *cmd = findSection(ctx, name: ".interp" , partition: partNo)) |
2335 | addHdr(PT_INTERP, cmd->getPhdrFlags())->add(cmd); |
2336 | |
2337 | // Add the headers. We will remove them if they don't fit. |
2338 | // In the other partitions the headers are ordinary sections, so they don't |
2339 | // need to be added here. |
2340 | if (isMain) { |
2341 | load = addHdr(PT_LOAD, flags); |
2342 | load->add(sec: ctx.out.elfHeader.get()); |
2343 | load->add(sec: ctx.out.programHeaders.get()); |
2344 | } |
2345 | } |
2346 | |
2347 | // PT_GNU_RELRO includes all sections that should be marked as |
2348 | // read-only by dynamic linker after processing relocations. |
2349 | // Current dynamic loaders only support one PT_GNU_RELRO PHDR, give |
2350 | // an error message if more than one PT_GNU_RELRO PHDR is required. |
2351 | auto relRo = std::make_unique<PhdrEntry>(args&: ctx, args: PT_GNU_RELRO, args: PF_R); |
2352 | bool inRelroPhdr = false; |
2353 | OutputSection *relroEnd = nullptr; |
2354 | for (OutputSection *sec : ctx.outputSections) { |
2355 | if (sec->partition != partNo || !needsPtLoad(sec)) |
2356 | continue; |
2357 | if (isRelroSection(ctx, sec)) { |
2358 | inRelroPhdr = true; |
2359 | if (!relroEnd) |
2360 | relRo->add(sec); |
2361 | else |
2362 | ErrAlways(ctx) << "section: " << sec->name |
2363 | << " is not contiguous with other relro" << " sections" ; |
2364 | } else if (inRelroPhdr) { |
2365 | inRelroPhdr = false; |
2366 | relroEnd = sec; |
2367 | } |
2368 | } |
2369 | relRo->p_align = 1; |
2370 | |
2371 | for (OutputSection *sec : ctx.outputSections) { |
2372 | if (!needsPtLoad(sec)) |
2373 | continue; |
2374 | |
2375 | // Normally, sections in partitions other than the current partition are |
2376 | // ignored. But partition number 255 is a special case: it contains the |
2377 | // partition end marker (.part.end). It needs to be added to the main |
2378 | // partition so that a segment is created for it in the main partition, |
2379 | // which will cause the dynamic loader to reserve space for the other |
2380 | // partitions. |
2381 | if (sec->partition != partNo) { |
2382 | if (isMain && sec->partition == 255) |
2383 | addHdr(PT_LOAD, computeFlags(ctx, flags: sec->getPhdrFlags()))->add(sec); |
2384 | continue; |
2385 | } |
2386 | |
2387 | // Segments are contiguous memory regions that has the same attributes |
2388 | // (e.g. executable or writable). There is one phdr for each segment. |
2389 | // Therefore, we need to create a new phdr when the next section has |
2390 | // incompatible flags or is loaded at a discontiguous address or memory |
2391 | // region using AT or AT> linker script command, respectively. |
2392 | // |
2393 | // As an exception, we don't create a separate load segment for the ELF |
2394 | // headers, even if the first "real" output has an AT or AT> attribute. |
2395 | // |
2396 | // In addition, NOBITS sections should only be placed at the end of a LOAD |
2397 | // segment (since it's represented as p_filesz < p_memsz). If we have a |
2398 | // not-NOBITS section after a NOBITS, we create a new LOAD for the latter |
2399 | // even if flags match, so as not to require actually writing the |
2400 | // supposed-to-be-NOBITS section to the output file. (However, we cannot do |
2401 | // so when hasSectionsCommand, since we cannot introduce the extra alignment |
2402 | // needed to create a new LOAD) |
2403 | uint64_t newFlags = computeFlags(ctx, flags: sec->getPhdrFlags()); |
2404 | uint64_t incompatible = flags ^ newFlags; |
2405 | if (!(newFlags & PF_W)) { |
2406 | // When --no-rosegment is specified, RO and RX sections are compatible. |
2407 | if (ctx.arg.singleRoRx) |
2408 | incompatible &= ~PF_X; |
2409 | // When --no-xosegment is specified (the default), XO and RX sections are |
2410 | // compatible. |
2411 | if (ctx.arg.singleXoRx) |
2412 | incompatible &= ~PF_R; |
2413 | } |
2414 | if (incompatible) |
2415 | load = nullptr; |
2416 | |
2417 | bool sameLMARegion = |
2418 | load && !sec->lmaExpr && sec->lmaRegion == load->firstSec->lmaRegion; |
2419 | if (load && sec != relroEnd && |
2420 | sec->memRegion == load->firstSec->memRegion && |
2421 | (sameLMARegion || load->lastSec == ctx.out.programHeaders.get()) && |
2422 | (ctx.script->hasSectionsCommand || sec->type == SHT_NOBITS || |
2423 | load->lastSec->type != SHT_NOBITS)) { |
2424 | load->p_flags |= newFlags; |
2425 | } else { |
2426 | load = addHdr(PT_LOAD, newFlags); |
2427 | flags = newFlags; |
2428 | } |
2429 | |
2430 | load->add(sec); |
2431 | } |
2432 | |
2433 | // Add a TLS segment if any. |
2434 | auto tlsHdr = std::make_unique<PhdrEntry>(args&: ctx, args: PT_TLS, args: PF_R); |
2435 | for (OutputSection *sec : ctx.outputSections) |
2436 | if (sec->partition == partNo && sec->flags & SHF_TLS) |
2437 | tlsHdr->add(sec); |
2438 | if (tlsHdr->firstSec) |
2439 | ret.push_back(Elt: std::move(tlsHdr)); |
2440 | |
2441 | // Add an entry for .dynamic. |
2442 | if (OutputSection *sec = part.dynamic->getParent()) |
2443 | addHdr(PT_DYNAMIC, sec->getPhdrFlags())->add(sec); |
2444 | |
2445 | if (relRo->firstSec) |
2446 | ret.push_back(Elt: std::move(relRo)); |
2447 | |
2448 | // PT_GNU_EH_FRAME is a special section pointing on .eh_frame_hdr. |
2449 | if (part.ehFrame->isNeeded() && part.ehFrameHdr && |
2450 | part.ehFrame->getParent() && part.ehFrameHdr->getParent()) |
2451 | addHdr(PT_GNU_EH_FRAME, part.ehFrameHdr->getParent()->getPhdrFlags()) |
2452 | ->add(part.ehFrameHdr->getParent()); |
2453 | |
2454 | if (ctx.arg.osabi == ELFOSABI_OPENBSD) { |
2455 | // PT_OPENBSD_MUTABLE makes the dynamic linker fill the segment with |
2456 | // zero data, like bss, but it can be treated differently. |
2457 | if (OutputSection *cmd = findSection(ctx, name: ".openbsd.mutable" , partition: partNo)) |
2458 | addHdr(PT_OPENBSD_MUTABLE, cmd->getPhdrFlags())->add(cmd); |
2459 | |
2460 | // PT_OPENBSD_RANDOMIZE makes the dynamic linker fill the segment |
2461 | // with random data. |
2462 | if (OutputSection *cmd = findSection(ctx, name: ".openbsd.randomdata" , partition: partNo)) |
2463 | addHdr(PT_OPENBSD_RANDOMIZE, cmd->getPhdrFlags())->add(cmd); |
2464 | |
2465 | // PT_OPENBSD_SYSCALLS makes the kernel and dynamic linker register |
2466 | // system call sites. |
2467 | if (OutputSection *cmd = findSection(ctx, name: ".openbsd.syscalls" , partition: partNo)) |
2468 | addHdr(PT_OPENBSD_SYSCALLS, cmd->getPhdrFlags())->add(cmd); |
2469 | } |
2470 | |
2471 | if (ctx.arg.zGnustack != GnuStackKind::None) { |
2472 | // PT_GNU_STACK is a special section to tell the loader to make the |
2473 | // pages for the stack non-executable. If you really want an executable |
2474 | // stack, you can pass -z execstack, but that's not recommended for |
2475 | // security reasons. |
2476 | unsigned perm = PF_R | PF_W; |
2477 | if (ctx.arg.zGnustack == GnuStackKind::Exec) |
2478 | perm |= PF_X; |
2479 | addHdr(PT_GNU_STACK, perm)->p_memsz = ctx.arg.zStackSize; |
2480 | } |
2481 | |
2482 | // PT_OPENBSD_NOBTCFI is an OpenBSD-specific header to mark that the |
2483 | // executable is expected to violate branch-target CFI checks. |
2484 | if (ctx.arg.zNoBtCfi) |
2485 | addHdr(PT_OPENBSD_NOBTCFI, PF_X); |
2486 | |
2487 | // PT_OPENBSD_WXNEEDED is a OpenBSD-specific header to mark the executable |
2488 | // is expected to perform W^X violations, such as calling mprotect(2) or |
2489 | // mmap(2) with PROT_WRITE | PROT_EXEC, which is prohibited by default on |
2490 | // OpenBSD. |
2491 | if (ctx.arg.zWxneeded) |
2492 | addHdr(PT_OPENBSD_WXNEEDED, PF_X); |
2493 | |
2494 | if (OutputSection *cmd = findSection(ctx, name: ".note.gnu.property" , partition: partNo)) |
2495 | addHdr(PT_GNU_PROPERTY, PF_R)->add(cmd); |
2496 | |
2497 | // Create one PT_NOTE per a group of contiguous SHT_NOTE sections with the |
2498 | // same alignment. |
2499 | PhdrEntry *note = nullptr; |
2500 | for (OutputSection *sec : ctx.outputSections) { |
2501 | if (sec->partition != partNo) |
2502 | continue; |
2503 | if (sec->type == SHT_NOTE && (sec->flags & SHF_ALLOC)) { |
2504 | if (!note || sec->lmaExpr || note->lastSec->addralign != sec->addralign) |
2505 | note = addHdr(PT_NOTE, PF_R); |
2506 | note->add(sec); |
2507 | } else { |
2508 | note = nullptr; |
2509 | } |
2510 | } |
2511 | return ret; |
2512 | } |
2513 | |
2514 | template <class ELFT> |
2515 | void Writer<ELFT>::addPhdrForSection(Partition &part, unsigned shType, |
2516 | unsigned pType, unsigned pFlags) { |
2517 | unsigned partNo = part.getNumber(ctx); |
2518 | auto i = llvm::find_if(ctx.outputSections, [=](OutputSection *cmd) { |
2519 | return cmd->partition == partNo && cmd->type == shType; |
2520 | }); |
2521 | if (i == ctx.outputSections.end()) |
2522 | return; |
2523 | |
2524 | auto entry = std::make_unique<PhdrEntry>(args&: ctx, args&: pType, args&: pFlags); |
2525 | entry->add(sec: *i); |
2526 | part.phdrs.push_back(Elt: std::move(entry)); |
2527 | } |
2528 | |
2529 | // Place the first section of each PT_LOAD to a different page (of maxPageSize). |
2530 | // This is achieved by assigning an alignment expression to addrExpr of each |
2531 | // such section. |
2532 | template <class ELFT> void Writer<ELFT>::fixSectionAlignments() { |
2533 | const PhdrEntry *prev; |
2534 | auto pageAlign = [&, &ctx = this->ctx](const PhdrEntry *p) { |
2535 | OutputSection *cmd = p->firstSec; |
2536 | if (!cmd) |
2537 | return; |
2538 | cmd->alignExpr = [align = cmd->addralign]() { return align; }; |
2539 | if (!cmd->addrExpr) { |
2540 | // Prefer advancing to align(dot, maxPageSize) + dot%maxPageSize to avoid |
2541 | // padding in the file contents. |
2542 | // |
2543 | // When -z separate-code is used we must not have any overlap in pages |
2544 | // between an executable segment and a non-executable segment. We align to |
2545 | // the next maximum page size boundary on transitions between executable |
2546 | // and non-executable segments. |
2547 | // |
2548 | // SHT_LLVM_PART_EHDR marks the start of a partition. The partition |
2549 | // sections will be extracted to a separate file. Align to the next |
2550 | // maximum page size boundary so that we can find the ELF header at the |
2551 | // start. We cannot benefit from overlapping p_offset ranges with the |
2552 | // previous segment anyway. |
2553 | if (ctx.arg.zSeparate == SeparateSegmentKind::Loadable || |
2554 | (ctx.arg.zSeparate == SeparateSegmentKind::Code && prev && |
2555 | (prev->p_flags & PF_X) != (p->p_flags & PF_X)) || |
2556 | cmd->type == SHT_LLVM_PART_EHDR) |
2557 | cmd->addrExpr = [&ctx = this->ctx] { |
2558 | return alignToPowerOf2(ctx.script->getDot(), ctx.arg.maxPageSize); |
2559 | }; |
2560 | // PT_TLS is at the start of the first RW PT_LOAD. If `p` includes PT_TLS, |
2561 | // it must be the RW. Align to p_align(PT_TLS) to make sure |
2562 | // p_vaddr(PT_LOAD)%p_align(PT_LOAD) = 0. Otherwise, if |
2563 | // sh_addralign(.tdata) < sh_addralign(.tbss), we will set p_align(PT_TLS) |
2564 | // to sh_addralign(.tbss), while p_vaddr(PT_TLS)=p_vaddr(PT_LOAD) may not |
2565 | // be congruent to 0 modulo p_align(PT_TLS). |
2566 | // |
2567 | // Technically this is not required, but as of 2019, some dynamic loaders |
2568 | // don't handle p_vaddr%p_align != 0 correctly, e.g. glibc (i386 and |
2569 | // x86-64) doesn't make runtime address congruent to p_vaddr modulo |
2570 | // p_align for dynamic TLS blocks (PR/24606), FreeBSD rtld has the same |
2571 | // bug, musl (TLS Variant 1 architectures) before 1.1.23 handled TLS |
2572 | // blocks correctly. We need to keep the workaround for a while. |
2573 | else if (ctx.tlsPhdr && ctx.tlsPhdr->firstSec == p->firstSec) |
2574 | cmd->addrExpr = [&ctx] { |
2575 | return alignToPowerOf2(ctx.script->getDot(), ctx.arg.maxPageSize) + |
2576 | alignToPowerOf2(ctx.script->getDot() % ctx.arg.maxPageSize, |
2577 | ctx.tlsPhdr->p_align); |
2578 | }; |
2579 | else |
2580 | cmd->addrExpr = [&ctx] { |
2581 | return alignToPowerOf2(ctx.script->getDot(), ctx.arg.maxPageSize) + |
2582 | ctx.script->getDot() % ctx.arg.maxPageSize; |
2583 | }; |
2584 | } |
2585 | }; |
2586 | |
2587 | for (Partition &part : ctx.partitions) { |
2588 | prev = nullptr; |
2589 | for (auto &p : part.phdrs) |
2590 | if (p->p_type == PT_LOAD && p->firstSec) { |
2591 | pageAlign(p.get()); |
2592 | prev = p.get(); |
2593 | } |
2594 | } |
2595 | } |
2596 | |
2597 | // Compute an in-file position for a given section. The file offset must be the |
2598 | // same with its virtual address modulo the page size, so that the loader can |
2599 | // load executables without any address adjustment. |
2600 | static uint64_t computeFileOffset(Ctx &ctx, OutputSection *os, uint64_t off) { |
2601 | // The first section in a PT_LOAD has to have congruent offset and address |
2602 | // modulo the maximum page size. |
2603 | if (os->ptLoad && os->ptLoad->firstSec == os) |
2604 | return alignTo(Value: off, Align: os->ptLoad->p_align, Skew: os->addr); |
2605 | |
2606 | // File offsets are not significant for .bss sections other than the first one |
2607 | // in a PT_LOAD/PT_TLS. By convention, we keep section offsets monotonically |
2608 | // increasing rather than setting to zero. |
2609 | if (os->type == SHT_NOBITS && (!ctx.tlsPhdr || ctx.tlsPhdr->firstSec != os)) |
2610 | return off; |
2611 | |
2612 | // If the section is not in a PT_LOAD, we just have to align it. |
2613 | if (!os->ptLoad) |
2614 | return alignToPowerOf2(Value: off, Align: os->addralign); |
2615 | |
2616 | // If two sections share the same PT_LOAD the file offset is calculated |
2617 | // using this formula: Off2 = Off1 + (VA2 - VA1). |
2618 | OutputSection *first = os->ptLoad->firstSec; |
2619 | return first->offset + os->addr - first->addr; |
2620 | } |
2621 | |
2622 | template <class ELFT> void Writer<ELFT>::assignFileOffsetsBinary() { |
2623 | // Compute the minimum LMA of all non-empty non-NOBITS sections as minAddr. |
2624 | auto needsOffset = [](OutputSection &sec) { |
2625 | return sec.type != SHT_NOBITS && (sec.flags & SHF_ALLOC) && sec.size > 0; |
2626 | }; |
2627 | uint64_t minAddr = UINT64_MAX; |
2628 | for (OutputSection *sec : ctx.outputSections) |
2629 | if (needsOffset(*sec)) { |
2630 | sec->offset = sec->getLMA(); |
2631 | minAddr = std::min(a: minAddr, b: sec->offset); |
2632 | } |
2633 | |
2634 | // Sections are laid out at LMA minus minAddr. |
2635 | fileSize = 0; |
2636 | for (OutputSection *sec : ctx.outputSections) |
2637 | if (needsOffset(*sec)) { |
2638 | sec->offset -= minAddr; |
2639 | fileSize = std::max(a: fileSize, b: sec->offset + sec->size); |
2640 | } |
2641 | } |
2642 | |
2643 | static std::string rangeToString(uint64_t addr, uint64_t len) { |
2644 | return "[0x" + utohexstr(X: addr) + ", 0x" + utohexstr(X: addr + len - 1) + "]" ; |
2645 | } |
2646 | |
2647 | // Assign file offsets to output sections. |
2648 | template <class ELFT> void Writer<ELFT>::assignFileOffsets() { |
2649 | ctx.out.programHeaders->offset = ctx.out.elfHeader->size; |
2650 | uint64_t off = ctx.out.elfHeader->size + ctx.out.programHeaders->size; |
2651 | |
2652 | PhdrEntry *lastRX = nullptr; |
2653 | for (Partition &part : ctx.partitions) |
2654 | for (auto &p : part.phdrs) |
2655 | if (p->p_type == PT_LOAD && (p->p_flags & PF_X)) |
2656 | lastRX = p.get(); |
2657 | |
2658 | // Layout SHF_ALLOC sections before non-SHF_ALLOC sections. A non-SHF_ALLOC |
2659 | // will not occupy file offsets contained by a PT_LOAD. |
2660 | for (OutputSection *sec : ctx.outputSections) { |
2661 | if (!(sec->flags & SHF_ALLOC)) |
2662 | continue; |
2663 | off = computeFileOffset(ctx, os: sec, off); |
2664 | sec->offset = off; |
2665 | if (sec->type != SHT_NOBITS) |
2666 | off += sec->size; |
2667 | |
2668 | // If this is a last section of the last executable segment and that |
2669 | // segment is the last loadable segment, align the offset of the |
2670 | // following section to avoid loading non-segments parts of the file. |
2671 | if (ctx.arg.zSeparate != SeparateSegmentKind::None && lastRX && |
2672 | lastRX->lastSec == sec) |
2673 | off = alignToPowerOf2(Value: off, Align: ctx.arg.maxPageSize); |
2674 | } |
2675 | for (OutputSection *osec : ctx.outputSections) { |
2676 | if (osec->flags & SHF_ALLOC) |
2677 | continue; |
2678 | osec->offset = alignToPowerOf2(Value: off, Align: osec->addralign); |
2679 | off = osec->offset + osec->size; |
2680 | } |
2681 | |
2682 | sectionHeaderOff = alignToPowerOf2(Value: off, Align: ctx.arg.wordsize); |
2683 | fileSize = |
2684 | sectionHeaderOff + (ctx.outputSections.size() + 1) * sizeof(Elf_Shdr); |
2685 | |
2686 | // Our logic assumes that sections have rising VA within the same segment. |
2687 | // With use of linker scripts it is possible to violate this rule and get file |
2688 | // offset overlaps or overflows. That should never happen with a valid script |
2689 | // which does not move the location counter backwards and usually scripts do |
2690 | // not do that. Unfortunately, there are apps in the wild, for example, Linux |
2691 | // kernel, which control segment distribution explicitly and move the counter |
2692 | // backwards, so we have to allow doing that to support linking them. We |
2693 | // perform non-critical checks for overlaps in checkSectionOverlap(), but here |
2694 | // we want to prevent file size overflows because it would crash the linker. |
2695 | for (OutputSection *sec : ctx.outputSections) { |
2696 | if (sec->type == SHT_NOBITS) |
2697 | continue; |
2698 | if ((sec->offset > fileSize) || (sec->offset + sec->size > fileSize)) |
2699 | ErrAlways(ctx) << "unable to place section " << sec->name |
2700 | << " at file offset " |
2701 | << rangeToString(addr: sec->offset, len: sec->size) |
2702 | << "; check your linker script for overflows" ; |
2703 | } |
2704 | } |
2705 | |
2706 | // Finalize the program headers. We call this function after we assign |
2707 | // file offsets and VAs to all sections. |
2708 | template <class ELFT> void Writer<ELFT>::setPhdrs(Partition &part) { |
2709 | for (std::unique_ptr<PhdrEntry> &p : part.phdrs) { |
2710 | OutputSection *first = p->firstSec; |
2711 | OutputSection *last = p->lastSec; |
2712 | |
2713 | // .ARM.exidx sections may not be within a single .ARM.exidx |
2714 | // output section. We always want to describe just the |
2715 | // SyntheticSection. |
2716 | if (part.armExidx && p->p_type == PT_ARM_EXIDX) { |
2717 | p->p_filesz = part.armExidx->getSize(); |
2718 | p->p_memsz = p->p_filesz; |
2719 | p->p_offset = first->offset + part.armExidx->outSecOff; |
2720 | p->p_vaddr = first->addr + part.armExidx->outSecOff; |
2721 | p->p_align = part.armExidx->addralign; |
2722 | if (part.elfHeader) |
2723 | p->p_offset -= part.elfHeader->getParent()->offset; |
2724 | |
2725 | if (!p->hasLMA) |
2726 | p->p_paddr = first->getLMA() + part.armExidx->outSecOff; |
2727 | return; |
2728 | } |
2729 | |
2730 | if (first) { |
2731 | p->p_filesz = last->offset - first->offset; |
2732 | if (last->type != SHT_NOBITS) |
2733 | p->p_filesz += last->size; |
2734 | |
2735 | p->p_memsz = last->addr + last->size - first->addr; |
2736 | p->p_offset = first->offset; |
2737 | p->p_vaddr = first->addr; |
2738 | |
2739 | // File offsets in partitions other than the main partition are relative |
2740 | // to the offset of the ELF headers. Perform that adjustment now. |
2741 | if (part.elfHeader) |
2742 | p->p_offset -= part.elfHeader->getParent()->offset; |
2743 | |
2744 | if (!p->hasLMA) |
2745 | p->p_paddr = first->getLMA(); |
2746 | } |
2747 | } |
2748 | } |
2749 | |
2750 | // A helper struct for checkSectionOverlap. |
2751 | namespace { |
2752 | struct SectionOffset { |
2753 | OutputSection *sec; |
2754 | uint64_t offset; |
2755 | }; |
2756 | } // namespace |
2757 | |
2758 | // Check whether sections overlap for a specific address range (file offsets, |
2759 | // load and virtual addresses). |
2760 | static void checkOverlap(Ctx &ctx, StringRef name, |
2761 | std::vector<SectionOffset> §ions, |
2762 | bool isVirtualAddr) { |
2763 | llvm::sort(C&: sections, Comp: [=](const SectionOffset &a, const SectionOffset &b) { |
2764 | return a.offset < b.offset; |
2765 | }); |
2766 | |
2767 | // Finding overlap is easy given a vector is sorted by start position. |
2768 | // If an element starts before the end of the previous element, they overlap. |
2769 | for (size_t i = 1, end = sections.size(); i < end; ++i) { |
2770 | SectionOffset a = sections[i - 1]; |
2771 | SectionOffset b = sections[i]; |
2772 | if (b.offset >= a.offset + a.sec->size) |
2773 | continue; |
2774 | |
2775 | // If both sections are in OVERLAY we allow the overlapping of virtual |
2776 | // addresses, because it is what OVERLAY was designed for. |
2777 | if (isVirtualAddr && a.sec->inOverlay && b.sec->inOverlay) |
2778 | continue; |
2779 | |
2780 | Err(ctx) << "section " << a.sec->name << " " << name |
2781 | << " range overlaps with " << b.sec->name << "\n>>> " |
2782 | << a.sec->name << " range is " |
2783 | << rangeToString(addr: a.offset, len: a.sec->size) << "\n>>> " << b.sec->name |
2784 | << " range is " << rangeToString(addr: b.offset, len: b.sec->size); |
2785 | } |
2786 | } |
2787 | |
2788 | // Check for overlapping sections and address overflows. |
2789 | // |
2790 | // In this function we check that none of the output sections have overlapping |
2791 | // file offsets. For SHF_ALLOC sections we also check that the load address |
2792 | // ranges and the virtual address ranges don't overlap |
2793 | template <class ELFT> void Writer<ELFT>::checkSections() { |
2794 | // First, check that section's VAs fit in available address space for target. |
2795 | for (OutputSection *os : ctx.outputSections) |
2796 | if ((os->addr + os->size < os->addr) || |
2797 | (!ELFT::Is64Bits && os->addr + os->size > uint64_t(UINT32_MAX) + 1)) |
2798 | Err(ctx) << "section " << os->name << " at 0x" |
2799 | << utohexstr(X: os->addr, LowerCase: true) << " of size 0x" |
2800 | << utohexstr(X: os->size, LowerCase: true) |
2801 | << " exceeds available address space" ; |
2802 | |
2803 | // Check for overlapping file offsets. In this case we need to skip any |
2804 | // section marked as SHT_NOBITS. These sections don't actually occupy space in |
2805 | // the file so Sec->Offset + Sec->Size can overlap with others. If --oformat |
2806 | // binary is specified only add SHF_ALLOC sections are added to the output |
2807 | // file so we skip any non-allocated sections in that case. |
2808 | std::vector<SectionOffset> fileOffs; |
2809 | for (OutputSection *sec : ctx.outputSections) |
2810 | if (sec->size > 0 && sec->type != SHT_NOBITS && |
2811 | (!ctx.arg.oFormatBinary || (sec->flags & SHF_ALLOC))) |
2812 | fileOffs.push_back(x: {.sec: sec, .offset: sec->offset}); |
2813 | checkOverlap(ctx, name: "file" , sections&: fileOffs, isVirtualAddr: false); |
2814 | |
2815 | // When linking with -r there is no need to check for overlapping virtual/load |
2816 | // addresses since those addresses will only be assigned when the final |
2817 | // executable/shared object is created. |
2818 | if (ctx.arg.relocatable) |
2819 | return; |
2820 | |
2821 | // Checking for overlapping virtual and load addresses only needs to take |
2822 | // into account SHF_ALLOC sections since others will not be loaded. |
2823 | // Furthermore, we also need to skip SHF_TLS sections since these will be |
2824 | // mapped to other addresses at runtime and can therefore have overlapping |
2825 | // ranges in the file. |
2826 | std::vector<SectionOffset> vmas; |
2827 | for (OutputSection *sec : ctx.outputSections) |
2828 | if (sec->size > 0 && (sec->flags & SHF_ALLOC) && !(sec->flags & SHF_TLS)) |
2829 | vmas.push_back(x: {.sec: sec, .offset: sec->addr}); |
2830 | checkOverlap(ctx, name: "virtual address" , sections&: vmas, isVirtualAddr: true); |
2831 | |
2832 | // Finally, check that the load addresses don't overlap. This will usually be |
2833 | // the same as the virtual addresses but can be different when using a linker |
2834 | // script with AT(). |
2835 | std::vector<SectionOffset> lmas; |
2836 | for (OutputSection *sec : ctx.outputSections) |
2837 | if (sec->size > 0 && (sec->flags & SHF_ALLOC) && !(sec->flags & SHF_TLS)) |
2838 | lmas.push_back(x: {.sec: sec, .offset: sec->getLMA()}); |
2839 | checkOverlap(ctx, name: "load address" , sections&: lmas, isVirtualAddr: false); |
2840 | } |
2841 | |
2842 | // The entry point address is chosen in the following ways. |
2843 | // |
2844 | // 1. the '-e' entry command-line option; |
2845 | // 2. the ENTRY(symbol) command in a linker control script; |
2846 | // 3. the value of the symbol _start, if present; |
2847 | // 4. the number represented by the entry symbol, if it is a number; |
2848 | // 5. the address 0. |
2849 | static uint64_t getEntryAddr(Ctx &ctx) { |
2850 | // Case 1, 2 or 3 |
2851 | if (Symbol *b = ctx.symtab->find(name: ctx.arg.entry)) |
2852 | return b->getVA(ctx); |
2853 | |
2854 | // Case 4 |
2855 | uint64_t addr; |
2856 | if (to_integer(S: ctx.arg.entry, Num&: addr)) |
2857 | return addr; |
2858 | |
2859 | // Case 5 |
2860 | if (ctx.arg.warnMissingEntry) |
2861 | Warn(ctx) << "cannot find entry symbol " << ctx.arg.entry |
2862 | << "; not setting start address" ; |
2863 | return 0; |
2864 | } |
2865 | |
2866 | static uint16_t getELFType(Ctx &ctx) { |
2867 | if (ctx.arg.isPic) |
2868 | return ET_DYN; |
2869 | if (ctx.arg.relocatable) |
2870 | return ET_REL; |
2871 | return ET_EXEC; |
2872 | } |
2873 | |
2874 | template <class ELFT> void Writer<ELFT>::() { |
2875 | writeEhdr<ELFT>(ctx, ctx.bufferStart, *ctx.mainPart); |
2876 | writePhdrs<ELFT>(ctx.bufferStart + sizeof(Elf_Ehdr), *ctx.mainPart); |
2877 | |
2878 | auto *eHdr = reinterpret_cast<Elf_Ehdr *>(ctx.bufferStart); |
2879 | eHdr->e_type = getELFType(ctx); |
2880 | eHdr->e_entry = getEntryAddr(ctx); |
2881 | |
2882 | // If -z nosectionheader is specified, omit the section header table. |
2883 | if (!ctx.in.shStrTab) |
2884 | return; |
2885 | eHdr->e_shoff = sectionHeaderOff; |
2886 | |
2887 | // Write the section header table. |
2888 | // |
2889 | // The ELF header can only store numbers up to SHN_LORESERVE in the e_shnum |
2890 | // and e_shstrndx fields. When the value of one of these fields exceeds |
2891 | // SHN_LORESERVE ELF requires us to put sentinel values in the ELF header and |
2892 | // use fields in the section header at index 0 to store |
2893 | // the value. The sentinel values and fields are: |
2894 | // e_shnum = 0, SHdrs[0].sh_size = number of sections. |
2895 | // e_shstrndx = SHN_XINDEX, SHdrs[0].sh_link = .shstrtab section index. |
2896 | auto *sHdrs = reinterpret_cast<Elf_Shdr *>(ctx.bufferStart + eHdr->e_shoff); |
2897 | size_t num = ctx.outputSections.size() + 1; |
2898 | if (num >= SHN_LORESERVE) |
2899 | sHdrs->sh_size = num; |
2900 | else |
2901 | eHdr->e_shnum = num; |
2902 | |
2903 | uint32_t strTabIndex = ctx.in.shStrTab->getParent()->sectionIndex; |
2904 | if (strTabIndex >= SHN_LORESERVE) { |
2905 | sHdrs->sh_link = strTabIndex; |
2906 | eHdr->e_shstrndx = SHN_XINDEX; |
2907 | } else { |
2908 | eHdr->e_shstrndx = strTabIndex; |
2909 | } |
2910 | |
2911 | for (OutputSection *sec : ctx.outputSections) |
2912 | sec->writeHeaderTo<ELFT>(++sHdrs); |
2913 | } |
2914 | |
2915 | // Open a result file. |
2916 | template <class ELFT> void Writer<ELFT>::openFile() { |
2917 | uint64_t maxSize = ctx.arg.is64 ? INT64_MAX : UINT32_MAX; |
2918 | if (fileSize != size_t(fileSize) || maxSize < fileSize) { |
2919 | std::string msg; |
2920 | raw_string_ostream s(msg); |
2921 | s << "output file too large: " << fileSize << " bytes\n" |
2922 | << "section sizes:\n" ; |
2923 | for (OutputSection *os : ctx.outputSections) |
2924 | s << os->name << ' ' << os->size << "\n" ; |
2925 | ErrAlways(ctx) << msg; |
2926 | return; |
2927 | } |
2928 | |
2929 | unlinkAsync(path: ctx.arg.outputFile); |
2930 | unsigned flags = 0; |
2931 | if (!ctx.arg.relocatable) |
2932 | flags |= FileOutputBuffer::F_executable; |
2933 | if (ctx.arg.mmapOutputFile) |
2934 | flags |= FileOutputBuffer::F_mmap; |
2935 | Expected<std::unique_ptr<FileOutputBuffer>> bufferOrErr = |
2936 | FileOutputBuffer::create(FilePath: ctx.arg.outputFile, Size: fileSize, Flags: flags); |
2937 | |
2938 | if (!bufferOrErr) { |
2939 | ErrAlways(ctx) << "failed to open " << ctx.arg.outputFile << ": " |
2940 | << bufferOrErr.takeError(); |
2941 | return; |
2942 | } |
2943 | buffer = std::move(*bufferOrErr); |
2944 | ctx.bufferStart = buffer->getBufferStart(); |
2945 | } |
2946 | |
2947 | template <class ELFT> void Writer<ELFT>::writeSectionsBinary() { |
2948 | parallel::TaskGroup tg; |
2949 | for (OutputSection *sec : ctx.outputSections) |
2950 | if (sec->flags & SHF_ALLOC) |
2951 | sec->writeTo<ELFT>(ctx, ctx.bufferStart + sec->offset, tg); |
2952 | } |
2953 | |
2954 | static void fillTrap(std::array<uint8_t, 4> trapInstr, uint8_t *i, |
2955 | uint8_t *end) { |
2956 | for (; i + 4 <= end; i += 4) |
2957 | memcpy(dest: i, src: trapInstr.data(), n: 4); |
2958 | } |
2959 | |
2960 | // Fill the last page of executable segments with trap instructions |
2961 | // instead of leaving them as zero. Even though it is not required by any |
2962 | // standard, it is in general a good thing to do for security reasons. |
2963 | // |
2964 | // We'll leave other pages in segments as-is because the rest will be |
2965 | // overwritten by output sections. |
2966 | template <class ELFT> void Writer<ELFT>::writeTrapInstr() { |
2967 | for (Partition &part : ctx.partitions) { |
2968 | // Fill the last page. |
2969 | for (std::unique_ptr<PhdrEntry> &p : part.phdrs) |
2970 | if (p->p_type == PT_LOAD && (p->p_flags & PF_X)) |
2971 | fillTrap( |
2972 | trapInstr: ctx.target->trapInstr, |
2973 | i: ctx.bufferStart + alignDown(Value: p->firstSec->offset + p->p_filesz, Align: 4), |
2974 | end: ctx.bufferStart + alignToPowerOf2(Value: p->firstSec->offset + p->p_filesz, |
2975 | Align: ctx.arg.maxPageSize)); |
2976 | |
2977 | // Round up the file size of the last segment to the page boundary iff it is |
2978 | // an executable segment to ensure that other tools don't accidentally |
2979 | // trim the instruction padding (e.g. when stripping the file). |
2980 | PhdrEntry *last = nullptr; |
2981 | for (std::unique_ptr<PhdrEntry> &p : part.phdrs) |
2982 | if (p->p_type == PT_LOAD) |
2983 | last = p.get(); |
2984 | |
2985 | if (last && (last->p_flags & PF_X)) { |
2986 | last->p_filesz = alignToPowerOf2(Value: last->p_filesz, Align: ctx.arg.maxPageSize); |
2987 | // p_memsz might be larger than the aligned p_filesz due to trailing BSS |
2988 | // sections. Don't decrease it. |
2989 | last->p_memsz = std::max(a: last->p_memsz, b: last->p_filesz); |
2990 | } |
2991 | } |
2992 | } |
2993 | |
2994 | // Write section contents to a mmap'ed file. |
2995 | template <class ELFT> void Writer<ELFT>::writeSections() { |
2996 | llvm::TimeTraceScope timeScope("Write sections" ); |
2997 | |
2998 | { |
2999 | // In -r or --emit-relocs mode, write the relocation sections first as in |
3000 | // ELf_Rel targets we might find out that we need to modify the relocated |
3001 | // section while doing it. |
3002 | parallel::TaskGroup tg; |
3003 | for (OutputSection *sec : ctx.outputSections) |
3004 | if (isStaticRelSecType(type: sec->type)) |
3005 | sec->writeTo<ELFT>(ctx, ctx.bufferStart + sec->offset, tg); |
3006 | } |
3007 | { |
3008 | parallel::TaskGroup tg; |
3009 | for (OutputSection *sec : ctx.outputSections) |
3010 | if (!isStaticRelSecType(type: sec->type)) |
3011 | sec->writeTo<ELFT>(ctx, ctx.bufferStart + sec->offset, tg); |
3012 | } |
3013 | |
3014 | // Finally, check that all dynamic relocation addends were written correctly. |
3015 | if (ctx.arg.checkDynamicRelocs && ctx.arg.writeAddends) { |
3016 | for (OutputSection *sec : ctx.outputSections) |
3017 | if (isStaticRelSecType(type: sec->type)) |
3018 | sec->checkDynRelAddends(ctx); |
3019 | } |
3020 | } |
3021 | |
3022 | // Computes a hash value of Data using a given hash function. |
3023 | // In order to utilize multiple cores, we first split data into 1MB |
3024 | // chunks, compute a hash for each chunk, and then compute a hash value |
3025 | // of the hash values. |
3026 | static void |
3027 | computeHash(llvm::MutableArrayRef<uint8_t> hashBuf, |
3028 | llvm::ArrayRef<uint8_t> data, |
3029 | std::function<void(uint8_t *dest, ArrayRef<uint8_t> arr)> hashFn) { |
3030 | std::vector<ArrayRef<uint8_t>> chunks = split(arr: data, chunkSize: 1024 * 1024); |
3031 | const size_t hashesSize = chunks.size() * hashBuf.size(); |
3032 | std::unique_ptr<uint8_t[]> hashes(new uint8_t[hashesSize]); |
3033 | |
3034 | // Compute hash values. |
3035 | parallelFor(Begin: 0, End: chunks.size(), Fn: [&](size_t i) { |
3036 | hashFn(hashes.get() + i * hashBuf.size(), chunks[i]); |
3037 | }); |
3038 | |
3039 | // Write to the final output buffer. |
3040 | hashFn(hashBuf.data(), ArrayRef(hashes.get(), hashesSize)); |
3041 | } |
3042 | |
3043 | template <class ELFT> void Writer<ELFT>::writeBuildId() { |
3044 | if (!ctx.mainPart->buildId || !ctx.mainPart->buildId->getParent()) |
3045 | return; |
3046 | |
3047 | if (ctx.arg.buildId == BuildIdKind::Hexstring) { |
3048 | for (Partition &part : ctx.partitions) |
3049 | part.buildId->writeBuildId(buf: ctx.arg.buildIdVector); |
3050 | return; |
3051 | } |
3052 | |
3053 | // Compute a hash of all sections of the output file. |
3054 | size_t hashSize = ctx.mainPart->buildId->hashSize; |
3055 | std::unique_ptr<uint8_t[]> buildId(new uint8_t[hashSize]); |
3056 | MutableArrayRef<uint8_t> output(buildId.get(), hashSize); |
3057 | llvm::ArrayRef<uint8_t> input{ctx.bufferStart, size_t(fileSize)}; |
3058 | |
3059 | // Fedora introduced build ID as "approximation of true uniqueness across all |
3060 | // binaries that might be used by overlapping sets of people". It does not |
3061 | // need some security goals that some hash algorithms strive to provide, e.g. |
3062 | // (second-)preimage and collision resistance. In practice people use 'md5' |
3063 | // and 'sha1' just for different lengths. Implement them with the more |
3064 | // efficient BLAKE3. |
3065 | switch (ctx.arg.buildId) { |
3066 | case BuildIdKind::Fast: |
3067 | computeHash(output, input, [](uint8_t *dest, ArrayRef<uint8_t> arr) { |
3068 | write64le(P: dest, V: xxh3_64bits(data: arr)); |
3069 | }); |
3070 | break; |
3071 | case BuildIdKind::Md5: |
3072 | computeHash(output, input, [&](uint8_t *dest, ArrayRef<uint8_t> arr) { |
3073 | memcpy(dest: dest, src: BLAKE3::hash<16>(Data: arr).data(), n: hashSize); |
3074 | }); |
3075 | break; |
3076 | case BuildIdKind::Sha1: |
3077 | computeHash(output, input, [&](uint8_t *dest, ArrayRef<uint8_t> arr) { |
3078 | memcpy(dest: dest, src: BLAKE3::hash<20>(Data: arr).data(), n: hashSize); |
3079 | }); |
3080 | break; |
3081 | case BuildIdKind::Uuid: |
3082 | if (auto ec = llvm::getRandomBytes(Buffer: buildId.get(), Size: hashSize)) |
3083 | ErrAlways(ctx) << "entropy source failure: " << ec.message(); |
3084 | break; |
3085 | default: |
3086 | llvm_unreachable("unknown BuildIdKind" ); |
3087 | } |
3088 | for (Partition &part : ctx.partitions) |
3089 | part.buildId->writeBuildId(buf: output); |
3090 | } |
3091 | |
3092 | template void elf::writeResult<ELF32LE>(Ctx &); |
3093 | template void elf::writeResult<ELF32BE>(Ctx &); |
3094 | template void elf::writeResult<ELF64LE>(Ctx &); |
3095 | template void elf::writeResult<ELF64BE>(Ctx &); |
3096 | |