1//===- MarkLive.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// This file implements --gc-sections, which is a feature to remove unused
10// sections from output. Unused sections are sections that are not reachable
11// from known GC-root symbols or sections. Naturally the feature is
12// implemented as a mark-sweep garbage collector.
13//
14// Here's how it works. Each InputSectionBase has a "Live" bit. The bit is off
15// by default. Starting with GC-root symbols or sections, markLive function
16// defined in this file visits all reachable sections to set their Live
17// bits. Writer will then ignore sections whose Live bits are off, so that
18// such sections are not included into output.
19//
20//===----------------------------------------------------------------------===//
21
22#include "MarkLive.h"
23#include "InputFiles.h"
24#include "InputSection.h"
25#include "LinkerScript.h"
26#include "SymbolTable.h"
27#include "Symbols.h"
28#include "SyntheticSections.h"
29#include "Target.h"
30#include "lld/Common/Strings.h"
31#include "llvm/ADT/DenseMapInfoVariant.h"
32#include "llvm/ADT/STLExtras.h"
33#include "llvm/Support/TimeProfiler.h"
34#include <variant>
35#include <vector>
36
37using namespace llvm;
38using namespace llvm::ELF;
39using namespace llvm::object;
40using namespace llvm::support::endian;
41using namespace lld;
42using namespace lld::elf;
43
44namespace {
45using SecOffset = std::pair<InputSectionBase *, unsigned>;
46
47// Something that can have an independent reason for being live.
48using LiveItem = std::variant<InputSectionBase *, Symbol *, SecOffset>;
49
50// The most proximate reason that something is live.
51struct LiveReason {
52 std::optional<LiveItem> item;
53 StringRef desc;
54};
55
56template <class ELFT, bool TrackWhyLive> class MarkLive {
57public:
58 MarkLive(Ctx &ctx, unsigned partition) : ctx(ctx), partition(partition) {}
59
60 void run();
61 void moveToMain();
62 void printWhyLive(Symbol *s) const;
63
64private:
65 void enqueue(InputSectionBase *sec, uint64_t offset, Symbol *sym,
66 LiveReason reason);
67 void markSymbol(Symbol *sym, StringRef reason);
68 void mark();
69
70 template <class RelTy>
71 void resolveReloc(InputSectionBase &sec, RelTy &rel, bool fromFDE);
72
73 template <class RelTy>
74 void scanEhFrameSection(EhInputSection &eh, ArrayRef<RelTy> rels);
75
76 Ctx &ctx;
77 // The index of the partition that we are currently processing.
78 unsigned partition;
79
80 // A list of sections to visit.
81 SmallVector<InputSection *, 0> queue;
82
83 // There are normally few input sections whose names are valid C
84 // identifiers, so we just store a SmallVector instead of a multimap.
85 DenseMap<StringRef, SmallVector<InputSectionBase *, 0>> cNamedSections;
86
87 // The most proximate reason that something is live. This forms a DAG between
88 // LiveItems. Acyclicality is maintained by only admitting the first
89 // discovered reason for each LiveItem; this captures the acyclic region of
90 // the liveness graph around the GC roots.
91 DenseMap<LiveItem, LiveReason> whyLive;
92};
93} // namespace
94
95template <class ELFT>
96static uint64_t getAddend(Ctx &ctx, InputSectionBase &sec,
97 const typename ELFT::Rel &rel) {
98 return ctx.target->getImplicitAddend(buf: sec.content().begin() + rel.r_offset,
99 type: rel.getType(ctx.arg.isMips64EL));
100}
101
102template <class ELFT>
103static uint64_t getAddend(Ctx &, InputSectionBase &sec,
104 const typename ELFT::Rela &rel) {
105 return rel.r_addend;
106}
107
108// Currently, we assume all input CREL relocations have an explicit addend.
109template <class ELFT>
110static uint64_t getAddend(Ctx &, InputSectionBase &sec,
111 const typename ELFT::Crel &rel) {
112 return rel.r_addend;
113}
114
115template <class ELFT, bool TrackWhyLive>
116template <class RelTy>
117void MarkLive<ELFT, TrackWhyLive>::resolveReloc(InputSectionBase &sec,
118 RelTy &rel, bool fromFDE) {
119 // If a symbol is referenced in a live section, it is used.
120 Symbol &sym = sec.file->getRelocTargetSym(rel);
121 sym.used = true;
122
123 LiveReason reason;
124 if (TrackWhyLive)
125 reason = {.item: SecOffset(&sec, rel.r_offset), .desc: "referenced by"};
126
127 if (auto *d = dyn_cast<Defined>(Val: &sym)) {
128 auto *relSec = dyn_cast_or_null<InputSectionBase>(Val: d->section);
129 if (!relSec)
130 return;
131
132 uint64_t offset = d->value;
133 if (d->isSection())
134 offset += getAddend<ELFT>(ctx, sec, rel);
135
136 // fromFDE being true means this is referenced by a FDE in a .eh_frame
137 // piece. The relocation points to the described function or to a LSDA. We
138 // only need to keep the LSDA live, so ignore anything that points to
139 // executable sections. If the LSDA is in a section group or has the
140 // SHF_LINK_ORDER flag, we ignore the relocation as well because (a) if the
141 // associated text section is live, the LSDA will be retained due to section
142 // group/SHF_LINK_ORDER rules (b) if the associated text section should be
143 // discarded, marking the LSDA will unnecessarily retain the text section.
144 if (!(fromFDE && ((relSec->flags & (SHF_EXECINSTR | SHF_LINK_ORDER)) ||
145 relSec->nextInSectionGroup))) {
146 Symbol *canonicalSym = d;
147 if (TrackWhyLive && d->isSection()) {
148 // This is expensive, so ideally this would be deferred until it's known
149 // whether this reference contributes to a printed whyLive chain, but
150 // that determination cannot be made without knowing the enclosing
151 // symbol.
152 if (Symbol *s = relSec->getEnclosingSymbol(offset))
153 canonicalSym = s;
154 else
155 canonicalSym = nullptr;
156 }
157 enqueue(sec: relSec, offset, sym: canonicalSym, reason);
158 }
159 return;
160 }
161
162 if (auto *ss = dyn_cast<SharedSymbol>(Val: &sym)) {
163 if (!ss->isWeak()) {
164 cast<SharedFile>(Val: ss->file)->isNeeded = true;
165 if (TrackWhyLive)
166 whyLive.try_emplace(Key: &sym, Args&: reason);
167 }
168 }
169
170 for (InputSectionBase *sec : cNamedSections.lookup(Val: sym.getName()))
171 enqueue(sec, /*offset=*/0, /*sym=*/nullptr, reason);
172}
173
174// The .eh_frame section is an unfortunate special case.
175// The section is divided in CIEs and FDEs and the relocations it can have are
176// * CIEs can refer to a personality function.
177// * FDEs can refer to a LSDA
178// * FDEs refer to the function they contain information about
179// The last kind of relocation cannot keep the referred section alive, or they
180// would keep everything alive in a common object file. In fact, each FDE is
181// alive if the section it refers to is alive.
182// To keep things simple, in here we just ignore the last relocation kind. The
183// other two keep the referred section alive.
184//
185// A possible improvement would be to fully process .eh_frame in the middle of
186// the gc pass. With that we would be able to also gc some sections holding
187// LSDAs and personality functions if we found that they were unused.
188template <class ELFT, bool TrackWhyLive>
189template <class RelTy>
190void MarkLive<ELFT, TrackWhyLive>::scanEhFrameSection(EhInputSection &eh,
191 ArrayRef<RelTy> rels) {
192 for (const EhSectionPiece &cie : eh.cies)
193 if (cie.firstRelocation != unsigned(-1))
194 resolveReloc(eh, rels[cie.firstRelocation], false);
195 for (const EhSectionPiece &fde : eh.fdes) {
196 size_t firstRelI = fde.firstRelocation;
197 if (firstRelI == (unsigned)-1)
198 continue;
199 uint64_t pieceEnd = fde.inputOff + fde.size;
200 for (size_t j = firstRelI, end2 = rels.size();
201 j < end2 && rels[j].r_offset < pieceEnd; ++j)
202 resolveReloc(eh, rels[j], true);
203 }
204}
205
206// Some sections are used directly by the loader, so they should never be
207// garbage-collected. This function returns true if a given section is such
208// section.
209static bool isReserved(InputSectionBase *sec) {
210 switch (sec->type) {
211 case SHT_FINI_ARRAY:
212 case SHT_INIT_ARRAY:
213 case SHT_PREINIT_ARRAY:
214 return true;
215 case SHT_NOTE:
216 // SHT_NOTE sections in a group are subject to garbage collection.
217 return !sec->nextInSectionGroup;
218 default:
219 // Support SHT_PROGBITS .init_array (https://golang.org/issue/50295) and
220 // .init_array.N (https://github.com/rust-lang/rust/issues/92181) for a
221 // while.
222 StringRef s = sec->name;
223 return s == ".init" || s == ".fini" || s.starts_with(Prefix: ".init_array") ||
224 s == ".jcr" || s.starts_with(Prefix: ".ctors") || s.starts_with(Prefix: ".dtors");
225 }
226}
227
228template <class ELFT, bool TrackWhyLive>
229void MarkLive<ELFT, TrackWhyLive>::enqueue(InputSectionBase *sec,
230 uint64_t offset, Symbol *sym,
231 LiveReason reason) {
232 // Usually, a whole section is marked as live or dead, but in mergeable
233 // (splittable) sections, each piece of data has independent liveness bit.
234 // So we explicitly tell it which offset is in use.
235 if (auto *ms = dyn_cast<MergeInputSection>(Val: sec))
236 ms->getSectionPiece(offset).live = true;
237
238 // Set Sec->Partition to the meet (i.e. the "minimum") of Partition and
239 // Sec->Partition in the following lattice: 1 < other < 0. If Sec->Partition
240 // doesn't change, we don't need to do anything.
241 if (sec->partition == 1 || sec->partition == partition)
242 return;
243 sec->partition = sec->partition ? 1 : partition;
244
245 if (TrackWhyLive) {
246 if (sym) {
247 // If a specific symbol is referenced, that keeps it live. The symbol then
248 // keeps its section live.
249 whyLive.try_emplace(Key: sym, Args&: reason);
250 whyLive.try_emplace(Key: sec, Args: LiveReason{.item: sym, .desc: "contained live symbol"});
251 } else {
252 // Otherwise, the reference generically keeps the section live.
253 whyLive.try_emplace(Key: sec, Args&: reason);
254 }
255 }
256
257 // Add input section to the queue.
258 if (InputSection *s = dyn_cast<InputSection>(Val: sec))
259 queue.push_back(Elt: s);
260}
261
262// Print the stack of reasons that the given symbol is live.
263template <class ELFT, bool TrackWhyLive>
264void MarkLive<ELFT, TrackWhyLive>::printWhyLive(Symbol *s) const {
265 // Skip dead symbols. A symbol is dead if it belongs to a dead section.
266 if (auto *d = dyn_cast<Defined>(Val: s)) {
267 auto *sec = dyn_cast_or_null<InputSectionBase>(Val: d->section);
268 if (sec && !sec->isLive())
269 return;
270 }
271
272 auto msg = Msg(ctx);
273
274 const auto printSymbol = [&](Symbol *s) {
275 msg << s->file << ":(" << s << ')';
276 };
277
278 msg << "live symbol: ";
279 printSymbol(s);
280
281 LiveItem cur = s;
282 while (true) {
283 auto it = whyLive.find(Val: cur);
284 LiveReason reason;
285 // If there is a specific reason this item is live...
286 if (it != whyLive.end()) {
287 reason = it->second;
288 } else {
289 // This item is live, but it has no tracked reason. It must be an
290 // unreferenced symbol in a live section or a symbol with no section.
291 InputSectionBase *sec = nullptr;
292 if (auto *d = dyn_cast<Defined>(Val: std::get<Symbol *>(v&: cur)))
293 sec = dyn_cast_or_null<InputSectionBase>(Val: d->section);
294 reason = sec ? LiveReason{.item: sec, .desc: "in live section"}
295 : LiveReason{.item: std::nullopt, .desc: "no section"};
296 }
297
298 if (!reason.item) {
299 msg << " (" << reason.desc << ')';
300 break;
301 }
302
303 msg << "\n>>> " << reason.desc << ": ";
304 // The reason may not yet have been resolved to a symbol; do so now.
305 if (std::holds_alternative<SecOffset>(v: *reason.item)) {
306 const auto &so = std::get<SecOffset>(v&: *reason.item);
307 InputSectionBase *sec = so.first;
308 Defined *sym = sec->getEnclosingSymbol(offset: so.second);
309 cur = sym ? LiveItem(sym) : LiveItem(sec);
310 } else {
311 cur = *reason.item;
312 }
313
314 if (std::holds_alternative<Symbol *>(v: cur))
315 printSymbol(std::get<Symbol *>(v&: cur));
316 else
317 msg << std::get<InputSectionBase *>(v&: cur);
318 }
319}
320
321template <class ELFT, bool TrackWhyLive>
322void MarkLive<ELFT, TrackWhyLive>::markSymbol(Symbol *sym, StringRef reason) {
323 if (auto *d = dyn_cast_or_null<Defined>(Val: sym))
324 if (auto *isec = dyn_cast_or_null<InputSectionBase>(Val: d->section))
325 enqueue(sec: isec, offset: d->value, sym, reason: {std::nullopt, reason});
326}
327
328// This is the main function of the garbage collector.
329// Starting from GC-root sections, this function visits all reachable
330// sections to set their "Live" bits.
331template <class ELFT, bool TrackWhyLive>
332void MarkLive<ELFT, TrackWhyLive>::run() {
333 // Add GC root symbols.
334
335 // Preserve externally-visible symbols if the symbols defined by this
336 // file can interpose other ELF file's symbols at runtime.
337 for (Symbol *sym : ctx.symtab->getSymbols())
338 if (sym->isExported && sym->partition == partition)
339 markSymbol(sym, reason: "externally visible symbol; may interpose");
340
341 // If this isn't the main partition, that's all that we need to preserve.
342 if (partition != 1) {
343 mark();
344 return;
345 }
346
347 markSymbol(sym: ctx.symtab->find(name: ctx.arg.entry), reason: "entry point");
348 markSymbol(sym: ctx.symtab->find(name: ctx.arg.init), reason: "initializer function");
349 markSymbol(sym: ctx.symtab->find(name: ctx.arg.fini), reason: "finalizer function");
350 for (StringRef s : ctx.arg.undefined)
351 markSymbol(sym: ctx.symtab->find(name: s), reason: "undefined command line flag");
352 for (StringRef s : ctx.script->referencedSymbols)
353 markSymbol(sym: ctx.symtab->find(name: s), reason: "referenced by linker script");
354 for (auto [symName, _] : ctx.symtab->cmseSymMap) {
355 markSymbol(sym: ctx.symtab->cmseSymMap[symName].sym, reason: "ARM CMSE symbol");
356 markSymbol(sym: ctx.symtab->cmseSymMap[symName].acleSeSym, reason: "ARM CMSE symbol");
357 }
358
359 // Mark .eh_frame sections as live because there are usually no relocations
360 // that point to .eh_frames. Otherwise, the garbage collector would drop
361 // all of them. We also want to preserve personality routines and LSDA
362 // referenced by .eh_frame sections, so we scan them for that here.
363 for (EhInputSection *eh : ctx.ehInputSections) {
364 const RelsOrRelas<ELFT> rels =
365 eh->template relsOrRelas<ELFT>(/*supportsCrel=*/false);
366 if (rels.areRelocsRel())
367 scanEhFrameSection(*eh, rels.rels);
368 else if (rels.relas.size())
369 scanEhFrameSection(*eh, rels.relas);
370 }
371 for (InputSectionBase *sec : ctx.inputSections) {
372 if (sec->flags & SHF_GNU_RETAIN) {
373 enqueue(sec, /*offset=*/0, /*sym=*/nullptr, reason: {std::nullopt, "retained"});
374 continue;
375 }
376 if (sec->flags & SHF_LINK_ORDER)
377 continue;
378
379 // Usually, non-SHF_ALLOC sections are not removed even if they are
380 // unreachable through relocations because reachability is not a good signal
381 // whether they are garbage or not (e.g. there is usually no section
382 // referring to a .comment section, but we want to keep it.) When a
383 // non-SHF_ALLOC section is retained, we also retain sections dependent on
384 // it.
385 //
386 // Note on SHF_LINK_ORDER: Such sections contain metadata and they
387 // have a reverse dependency on the InputSection they are linked with.
388 // We are able to garbage collect them.
389 //
390 // Note on SHF_REL{,A}: Such sections reach here only when -r
391 // or --emit-reloc were given. And they are subject of garbage
392 // collection because, if we remove a text section, we also
393 // remove its relocation section.
394 //
395 // Note on nextInSectionGroup: The ELF spec says that group sections are
396 // included or omitted as a unit. We take the interpretation that:
397 //
398 // - Group members (nextInSectionGroup != nullptr) are subject to garbage
399 // collection.
400 // - Groups members are retained or discarded as a unit.
401 if (!(sec->flags & SHF_ALLOC)) {
402 if (!isStaticRelSecType(type: sec->type) && !sec->nextInSectionGroup) {
403 sec->markLive();
404 for (InputSection *isec : sec->dependentSections)
405 isec->markLive();
406 }
407 }
408
409 // Preserve special sections and those which are specified in linker
410 // script KEEP command.
411 if (isReserved(sec)) {
412 enqueue(sec, /*offset=*/0, /*sym=*/nullptr, reason: {std::nullopt, "reserved"});
413 } else if (ctx.script->shouldKeep(s: sec)) {
414 enqueue(sec, /*offset=*/0, /*sym=*/nullptr,
415 reason: {std::nullopt, "KEEP in linker script"});
416 } else if ((!ctx.arg.zStartStopGC || sec->name.starts_with(Prefix: "__libc_")) &&
417 isValidCIdentifier(s: sec->name)) {
418 // As a workaround for glibc libc.a before 2.34
419 // (https://sourceware.org/PR27492), retain __libc_atexit and similar
420 // sections regardless of zStartStopGC.
421 cNamedSections[ctx.saver.save(S: "__start_" + sec->name)].push_back(Elt: sec);
422 cNamedSections[ctx.saver.save(S: "__stop_" + sec->name)].push_back(Elt: sec);
423 }
424 }
425
426 mark();
427
428 if (TrackWhyLive) {
429 const auto handleSym = [&](Symbol *sym) {
430 if (llvm::any_of(ctx.arg.whyLive, [sym](const llvm::GlobPattern &pat) {
431 return pat.match(S: sym->getName());
432 }))
433 printWhyLive(s: sym);
434 };
435
436 for (Symbol *sym : ctx.symtab->getSymbols())
437 handleSym(sym);
438 for (ELFFileBase *file : ctx.objectFiles)
439 for (Symbol *sym : file->getSymbols())
440 if (sym->isLocal())
441 handleSym(sym);
442 }
443}
444
445template <class ELFT, bool TrackWhyLive>
446void MarkLive<ELFT, TrackWhyLive>::mark() {
447 // Mark all reachable sections.
448 while (!queue.empty()) {
449 InputSectionBase &sec = *queue.pop_back_val();
450
451 const RelsOrRelas<ELFT> rels = sec.template relsOrRelas<ELFT>();
452 for (const typename ELFT::Rel &rel : rels.rels)
453 resolveReloc(sec, rel, false);
454 for (const typename ELFT::Rela &rel : rels.relas)
455 resolveReloc(sec, rel, false);
456 for (const typename ELFT::Crel &rel : rels.crels)
457 resolveReloc(sec, rel, false);
458
459 for (InputSectionBase *isec : sec.dependentSections)
460 enqueue(sec: isec, /*offset=*/0, /*sym=*/nullptr,
461 reason: {&sec, "depended on by section"});
462
463 // Mark the next group member.
464 if (sec.nextInSectionGroup)
465 enqueue(sec: sec.nextInSectionGroup, /*offset=*/0, /*sym=*/nullptr,
466 reason: {&sec, "in section group with"});
467 }
468}
469
470// Move the sections for some symbols to the main partition, specifically ifuncs
471// (because they can result in an IRELATIVE being added to the main partition's
472// GOT, which means that the ifunc must be available when the main partition is
473// loaded) and TLS symbols (because we only know how to correctly process TLS
474// relocations for the main partition).
475//
476// We also need to move sections whose names are C identifiers that are referred
477// to from __start_/__stop_ symbols because there will only be one set of
478// symbols for the whole program.
479template <class ELFT, bool TrackWhyLive>
480void MarkLive<ELFT, TrackWhyLive>::moveToMain() {
481 for (ELFFileBase *file : ctx.objectFiles)
482 for (Symbol *s : file->getSymbols())
483 if (auto *d = dyn_cast<Defined>(Val: s))
484 if ((d->type == STT_GNU_IFUNC || d->type == STT_TLS) && d->section &&
485 d->section->isLive())
486 markSymbol(sym: s, /*reason=*/{});
487
488 for (InputSectionBase *sec : ctx.inputSections) {
489 if (!sec->isLive() || !isValidCIdentifier(s: sec->name))
490 continue;
491 if (ctx.symtab->find(name: ("__start_" + sec->name).str()) ||
492 ctx.symtab->find(name: ("__stop_" + sec->name).str()))
493 enqueue(sec, /*offset=*/0, /*sym=*/nullptr, /*reason=*/{});
494 }
495
496 mark();
497}
498
499// Before calling this function, Live bits are off for all
500// input sections. This function make some or all of them on
501// so that they are emitted to the output file.
502template <class ELFT> void elf::markLive(Ctx &ctx) {
503 llvm::TimeTraceScope timeScope("markLive");
504 // If --gc-sections is not given, retain all input sections.
505 if (!ctx.arg.gcSections) {
506 // If a DSO defines a symbol referenced in a regular object, it is needed.
507 for (Symbol *sym : ctx.symtab->getSymbols())
508 if (auto *s = dyn_cast<SharedSymbol>(Val: sym))
509 if (s->isUsedInRegularObj && !s->isWeak())
510 cast<SharedFile>(Val: s->file)->isNeeded = true;
511 return;
512 }
513
514 for (InputSectionBase *sec : ctx.inputSections)
515 sec->markDead();
516
517 // Follow the graph to mark all live sections.
518 for (unsigned i = 1, e = ctx.partitions.size(); i <= e; ++i)
519 if (ctx.arg.whyLive.empty())
520 MarkLive<ELFT, false>(ctx, i).run();
521 else
522 MarkLive<ELFT, true>(ctx, i).run();
523
524 // If we have multiple partitions, some sections need to live in the main
525 // partition even if they were allocated to a loadable partition. Move them
526 // there now.
527 if (ctx.partitions.size() != 1)
528 MarkLive<ELFT, false>(ctx, 1).moveToMain();
529
530 // Report garbage-collected sections.
531 if (ctx.arg.printGcSections)
532 for (InputSectionBase *sec : ctx.inputSections)
533 if (!sec->isLive())
534 Msg(ctx) << "removing unused section " << sec;
535}
536
537template void elf::markLive<ELF32LE>(Ctx &);
538template void elf::markLive<ELF32BE>(Ctx &);
539template void elf::markLive<ELF64LE>(Ctx &);
540template void elf::markLive<ELF64BE>(Ctx &);
541