1 //===- Relocations.cpp ----------------------------------------------------===//
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains platform-independent functions to process relocations.
11 // I'll describe the overview of this file here.
13 // Simple relocations are easy to handle for the linker. For example,
14 // for R_X86_64_PC64 relocs, the linker just has to fix up locations
15 // with the relative offsets to the target symbols. It would just be
16 // reading records from relocation sections and applying them to output.
18 // But not all relocations are that easy to handle. For example, for
19 // R_386_GOTOFF relocs, the linker has to create new GOT entries for
20 // symbols if they don't exist, and fix up locations with GOT entry
21 // offsets from the beginning of GOT section. So there is more than
22 // fixing addresses in relocation processing.
24 // ELF defines a large number of complex relocations.
26 // The functions in this file analyze relocations and do whatever needs
27 // to be done. It includes, but not limited to, the following.
29 // - create GOT/PLT entries
30 // - create new relocations in .dynsym to let the dynamic linker resolve
31 // them at runtime (since ELF supports dynamic linking, not all
32 // relocations can be resolved at link-time)
33 // - create COPY relocs and reserve space in .bss
34 // - replace expensive relocs (in terms of runtime cost) with cheap ones
35 // - error out infeasible combinations such as PIC and non-relative relocs
37 // Note that the functions in this file don't actually apply relocations
38 // because it doesn't know about the output file nor the output file buffer.
39 // It instead stores Relocation objects to InputSection's Relocations
40 // vector to let it apply later in InputSection::writeTo.
42 //===----------------------------------------------------------------------===//
44 #include "Relocations.h"
46 #include "LinkerScript.h"
47 #include "OutputSections.h"
48 #include "SymbolTable.h"
50 #include "SyntheticSections.h"
53 #include "lld/Common/Memory.h"
54 #include "lld/Common/Strings.h"
55 #include "llvm/ADT/SmallSet.h"
56 #include "llvm/Support/Endian.h"
57 #include "llvm/Support/raw_ostream.h"
61 using namespace llvm::ELF;
62 using namespace llvm::object;
63 using namespace llvm::support::endian;
66 using namespace lld::elf;
68 // Construct a message in the following format.
70 // >>> defined in /home/alice/src/foo.o
71 // >>> referenced by bar.c:12 (/home/alice/src/bar.c:12)
72 // >>> /home/alice/src/bar.o:(.text+0x1)
73 static std::string getLocation(InputSectionBase &S, const Symbol &Sym,
76 "\n>>> defined in " + toString(Sym.File) + "\n>>> referenced by ";
77 std::string Src = S.getSrcMsg(Sym, Off);
79 Msg += Src + "\n>>> ";
80 return Msg + S.getObjMsg(Off);
83 // This function is similar to the `handleTlsRelocation`. MIPS does not
84 // support any relaxations for TLS relocations so by factoring out MIPS
85 // handling in to the separate function we can simplify the code and do not
86 // pollute other `handleTlsRelocation` by MIPS `ifs` statements.
87 // Mips has a custom MipsGotSection that handles the writing of GOT entries
88 // without dynamic relocations.
89 static unsigned handleMipsTlsRelocation(RelType Type, Symbol &Sym,
90 InputSectionBase &C, uint64_t Offset,
91 int64_t Addend, RelExpr Expr) {
92 if (Expr == R_MIPS_TLSLD) {
93 InX::MipsGot->addTlsIndex(*C.File);
94 C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
97 if (Expr == R_MIPS_TLSGD) {
98 InX::MipsGot->addDynTlsEntry(*C.File, Sym);
99 C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
105 // This function is similar to the `handleMipsTlsRelocation`. ARM also does not
106 // support any relaxations for TLS relocations. ARM is logically similar to Mips
107 // in how it handles TLS, but Mips uses its own custom GOT which handles some
108 // of the cases that ARM uses GOT relocations for.
110 // We look for TLS global dynamic and local dynamic relocations, these may
111 // require the generation of a pair of GOT entries that have associated
112 // dynamic relocations. When the results of the dynamic relocations can be
113 // resolved at static link time we do so. This is necessary for static linking
114 // as there will be no dynamic loader to resolve them at load-time.
116 // The pair of GOT entries created are of the form
117 // GOT[e0] Module Index (Used to find pointer to TLS block at run-time)
118 // GOT[e1] Offset of symbol in TLS block
119 template <class ELFT>
120 static unsigned handleARMTlsRelocation(RelType Type, Symbol &Sym,
121 InputSectionBase &C, uint64_t Offset,
122 int64_t Addend, RelExpr Expr) {
123 // The Dynamic TLS Module Index Relocation for a symbol defined in an
124 // executable is always 1. If the target Symbol is not preemptible then
125 // we know the offset into the TLS block at static link time.
126 bool NeedDynId = Sym.IsPreemptible || Config->Shared;
127 bool NeedDynOff = Sym.IsPreemptible;
129 auto AddTlsReloc = [&](uint64_t Off, RelType Type, Symbol *Dest, bool Dyn) {
131 InX::RelaDyn->addReloc(Type, InX::Got, Off, Dest);
133 InX::Got->Relocations.push_back({R_ABS, Type, Off, 0, Dest});
136 // Local Dynamic is for access to module local TLS variables, while still
137 // being suitable for being dynamically loaded via dlopen.
138 // GOT[e0] is the module index, with a special value of 0 for the current
139 // module. GOT[e1] is unused. There only needs to be one module index entry.
140 if (Expr == R_TLSLD_PC && InX::Got->addTlsIndex()) {
141 AddTlsReloc(InX::Got->getTlsIndexOff(), Target->TlsModuleIndexRel,
142 NeedDynId ? nullptr : &Sym, NeedDynId);
143 C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
147 // Global Dynamic is the most general purpose access model. When we know
148 // the module index and offset of symbol in TLS block we can fill these in
149 // using static GOT relocations.
150 if (Expr == R_TLSGD_PC) {
151 if (InX::Got->addDynTlsEntry(Sym)) {
152 uint64_t Off = InX::Got->getGlobalDynOffset(Sym);
153 AddTlsReloc(Off, Target->TlsModuleIndexRel, &Sym, NeedDynId);
154 AddTlsReloc(Off + Config->Wordsize, Target->TlsOffsetRel, &Sym,
157 C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
163 // Returns the number of relocations processed.
164 template <class ELFT>
166 handleTlsRelocation(RelType Type, Symbol &Sym, InputSectionBase &C,
167 typename ELFT::uint Offset, int64_t Addend, RelExpr Expr) {
168 if (!(C.Flags & SHF_ALLOC))
174 if (Config->EMachine == EM_ARM)
175 return handleARMTlsRelocation<ELFT>(Type, Sym, C, Offset, Addend, Expr);
176 if (Config->EMachine == EM_MIPS)
177 return handleMipsTlsRelocation(Type, Sym, C, Offset, Addend, Expr);
179 if (isRelExprOneOf<R_TLSDESC, R_TLSDESC_PAGE, R_TLSDESC_CALL>(Expr) &&
181 if (InX::Got->addDynTlsEntry(Sym)) {
182 uint64_t Off = InX::Got->getGlobalDynOffset(Sym);
183 InX::RelaDyn->addReloc(
184 {Target->TlsDescRel, InX::Got, Off, !Sym.IsPreemptible, &Sym, 0});
186 if (Expr != R_TLSDESC_CALL)
187 C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
191 if (isRelExprOneOf<R_TLSLD_GOT, R_TLSLD_GOT_FROM_END, R_TLSLD_PC,
192 R_TLSLD_HINT>(Expr)) {
193 // Local-Dynamic relocs can be relaxed to Local-Exec.
194 if (!Config->Shared) {
195 C.Relocations.push_back(
196 {Target->adjustRelaxExpr(Type, nullptr, R_RELAX_TLS_LD_TO_LE), Type,
197 Offset, Addend, &Sym});
198 return Target->TlsGdRelaxSkip;
200 if (Expr == R_TLSLD_HINT)
202 if (InX::Got->addTlsIndex())
203 InX::RelaDyn->addReloc(Target->TlsModuleIndexRel, InX::Got,
204 InX::Got->getTlsIndexOff(), nullptr);
205 C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
209 // Local-Dynamic relocs can be relaxed to Local-Exec.
210 if (Expr == R_ABS && !Config->Shared) {
211 C.Relocations.push_back(
212 {Target->adjustRelaxExpr(Type, nullptr, R_RELAX_TLS_LD_TO_LE), Type,
213 Offset, Addend, &Sym});
217 // Local-Dynamic sequence where offset of tls variable relative to dynamic
218 // thread pointer is stored in the got.
219 if (Expr == R_TLSLD_GOT_OFF) {
220 // Local-Dynamic relocs can be relaxed to local-exec
221 if (!Config->Shared) {
222 C.Relocations.push_back({R_RELAX_TLS_LD_TO_LE, Type, Offset, Addend, &Sym});
225 if (!Sym.isInGot()) {
226 InX::Got->addEntry(Sym);
227 uint64_t Off = Sym.getGotOffset();
228 InX::Got->Relocations.push_back({R_ABS, Target->TlsOffsetRel, Off, 0, &Sym});
230 C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
234 if (isRelExprOneOf<R_TLSDESC, R_TLSDESC_PAGE, R_TLSDESC_CALL, R_TLSGD_GOT,
235 R_TLSGD_GOT_FROM_END, R_TLSGD_PC>(Expr)) {
236 if (Config->Shared) {
237 if (InX::Got->addDynTlsEntry(Sym)) {
238 uint64_t Off = InX::Got->getGlobalDynOffset(Sym);
239 InX::RelaDyn->addReloc(Target->TlsModuleIndexRel, InX::Got, Off, &Sym);
241 // If the symbol is preemptible we need the dynamic linker to write
243 uint64_t OffsetOff = Off + Config->Wordsize;
244 if (Sym.IsPreemptible)
245 InX::RelaDyn->addReloc(Target->TlsOffsetRel, InX::Got, OffsetOff,
248 InX::Got->Relocations.push_back(
249 {R_ABS, Target->TlsOffsetRel, OffsetOff, 0, &Sym});
251 C.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
255 // Global-Dynamic relocs can be relaxed to Initial-Exec or Local-Exec
256 // depending on the symbol being locally defined or not.
257 if (Sym.IsPreemptible) {
258 C.Relocations.push_back(
259 {Target->adjustRelaxExpr(Type, nullptr, R_RELAX_TLS_GD_TO_IE), Type,
260 Offset, Addend, &Sym});
261 if (!Sym.isInGot()) {
262 InX::Got->addEntry(Sym);
263 InX::RelaDyn->addReloc(Target->TlsGotRel, InX::Got, Sym.getGotOffset(),
267 C.Relocations.push_back(
268 {Target->adjustRelaxExpr(Type, nullptr, R_RELAX_TLS_GD_TO_LE), Type,
269 Offset, Addend, &Sym});
271 return Target->TlsGdRelaxSkip;
274 // Initial-Exec relocs can be relaxed to Local-Exec if the symbol is locally
276 if (isRelExprOneOf<R_GOT, R_GOT_FROM_END, R_GOT_PC, R_GOT_PAGE_PC>(Expr) &&
277 !Config->Shared && !Sym.IsPreemptible) {
278 C.Relocations.push_back({R_RELAX_TLS_IE_TO_LE, Type, Offset, Addend, &Sym});
282 if (Expr == R_TLSDESC_CALL)
287 static RelType getMipsPairType(RelType Type, bool IsLocal) {
292 // In case of global symbol, the R_MIPS_GOT16 relocation does not
293 // have a pair. Each global symbol has a unique entry in the GOT
294 // and a corresponding instruction with help of the R_MIPS_GOT16
295 // relocation loads an address of the symbol. In case of local
296 // symbol, the R_MIPS_GOT16 relocation creates a GOT entry to hold
297 // the high 16 bits of the symbol's value. A paired R_MIPS_LO16
298 // relocations handle low 16 bits of the address. That allows
299 // to allocate only one GOT entry for every 64 KBytes of local data.
300 return IsLocal ? R_MIPS_LO16 : R_MIPS_NONE;
301 case R_MICROMIPS_GOT16:
302 return IsLocal ? R_MICROMIPS_LO16 : R_MIPS_NONE;
304 return R_MIPS_PCLO16;
305 case R_MICROMIPS_HI16:
306 return R_MICROMIPS_LO16;
312 // True if non-preemptable symbol always has the same value regardless of where
313 // the DSO is loaded.
314 static bool isAbsolute(const Symbol &Sym) {
315 if (Sym.isUndefWeak())
317 if (const auto *DR = dyn_cast<Defined>(&Sym))
318 return DR->Section == nullptr; // Absolute symbol.
322 static bool isAbsoluteValue(const Symbol &Sym) {
323 return isAbsolute(Sym) || Sym.isTls();
326 // Returns true if Expr refers a PLT entry.
327 static bool needsPlt(RelExpr Expr) {
328 return isRelExprOneOf<R_PLT_PC, R_PPC_CALL_PLT, R_PLT, R_PLT_PAGE_PC>(Expr);
331 // Returns true if Expr refers a GOT entry. Note that this function
332 // returns false for TLS variables even though they need GOT, because
333 // TLS variables uses GOT differently than the regular variables.
334 static bool needsGot(RelExpr Expr) {
335 return isRelExprOneOf<R_GOT, R_GOT_OFF, R_MIPS_GOT_LOCAL_PAGE, R_MIPS_GOT_OFF,
336 R_MIPS_GOT_OFF32, R_GOT_PAGE_PC, R_GOT_PC,
337 R_GOT_FROM_END>(Expr);
340 // True if this expression is of the form Sym - X, where X is a position in the
341 // file (PC, or GOT for example).
342 static bool isRelExpr(RelExpr Expr) {
343 return isRelExprOneOf<R_PC, R_GOTREL, R_GOTREL_FROM_END, R_MIPS_GOTREL,
344 R_PPC_CALL, R_PPC_CALL_PLT, R_PAGE_PC,
345 R_RELAX_GOT_PC>(Expr);
348 // Returns true if a given relocation can be computed at link-time.
350 // For instance, we know the offset from a relocation to its target at
351 // link-time if the relocation is PC-relative and refers a
352 // non-interposable function in the same executable. This function
353 // will return true for such relocation.
355 // If this function returns false, that means we need to emit a
356 // dynamic relocation so that the relocation will be fixed at load-time.
357 static bool isStaticLinkTimeConstant(RelExpr E, RelType Type, const Symbol &Sym,
358 InputSectionBase &S, uint64_t RelOff) {
359 // These expressions always compute a constant
361 R_GOT_FROM_END, R_GOT_OFF, R_TLSLD_GOT_OFF, R_MIPS_GOT_LOCAL_PAGE,
362 R_MIPS_GOTREL, R_MIPS_GOT_OFF, R_MIPS_GOT_OFF32, R_MIPS_GOT_GP_PC,
363 R_MIPS_TLSGD, R_GOT_PAGE_PC, R_GOT_PC, R_GOTONLY_PC,
364 R_GOTONLY_PC_FROM_END, R_PLT_PC, R_TLSGD_GOT, R_TLSGD_GOT_FROM_END,
365 R_TLSGD_PC, R_PPC_CALL_PLT, R_TLSDESC_CALL, R_TLSDESC_PAGE, R_HINT,
369 // The computation involves output from the ifunc resolver.
370 if (Sym.isGnuIFunc() && Config->ZIfuncnoplt)
373 // These never do, except if the entire file is position dependent or if
374 // only the low bits are used.
375 if (E == R_GOT || E == R_PLT || E == R_TLSDESC)
376 return Target->usesOnlyLowPageBits(Type) || !Config->Pic;
378 if (Sym.IsPreemptible)
383 // The size of a non preemptible symbol is a constant.
387 // For the target and the relocation, we want to know if they are
388 // absolute or relative.
389 bool AbsVal = isAbsoluteValue(Sym);
390 bool RelE = isRelExpr(E);
395 if (!AbsVal && !RelE)
396 return Target->usesOnlyLowPageBits(Type);
398 // Relative relocation to an absolute value. This is normally unrepresentable,
399 // but if the relocation refers to a weak undefined symbol, we allow it to
400 // resolve to the image base. This is a little strange, but it allows us to
401 // link function calls to such symbols. Normally such a call will be guarded
402 // with a comparison, which will load a zero from the GOT.
403 // Another special case is MIPS _gp_disp symbol which represents offset
404 // between start of a function and '_gp' value and defined as absolute just
405 // to simplify the code.
406 assert(AbsVal && RelE);
407 if (Sym.isUndefWeak())
410 error("relocation " + toString(Type) + " cannot refer to absolute symbol: " +
411 toString(Sym) + getLocation(S, Sym, RelOff));
415 static RelExpr toPlt(RelExpr Expr) {
418 return R_PPC_CALL_PLT;
422 return R_PLT_PAGE_PC;
430 static RelExpr fromPlt(RelExpr Expr) {
431 // We decided not to use a plt. Optimize a reference to the plt to a
432 // reference to the symbol itself.
445 // Returns true if a given shared symbol is in a read-only segment in a DSO.
446 template <class ELFT> static bool isReadOnly(SharedSymbol &SS) {
447 typedef typename ELFT::Phdr Elf_Phdr;
449 // Determine if the symbol is read-only by scanning the DSO's program headers.
450 const SharedFile<ELFT> &File = SS.getFile<ELFT>();
451 for (const Elf_Phdr &Phdr : check(File.getObj().program_headers()))
452 if ((Phdr.p_type == ELF::PT_LOAD || Phdr.p_type == ELF::PT_GNU_RELRO) &&
453 !(Phdr.p_flags & ELF::PF_W) && SS.Value >= Phdr.p_vaddr &&
454 SS.Value < Phdr.p_vaddr + Phdr.p_memsz)
459 // Returns symbols at the same offset as a given symbol, including SS itself.
461 // If two or more symbols are at the same offset, and at least one of
462 // them are copied by a copy relocation, all of them need to be copied.
463 // Otherwise, they would refer to different places at runtime.
464 template <class ELFT>
465 static SmallSet<SharedSymbol *, 4> getSymbolsAt(SharedSymbol &SS) {
466 typedef typename ELFT::Sym Elf_Sym;
468 SharedFile<ELFT> &File = SS.getFile<ELFT>();
470 SmallSet<SharedSymbol *, 4> Ret;
471 for (const Elf_Sym &S : File.getGlobalELFSyms()) {
472 if (S.st_shndx == SHN_UNDEF || S.st_shndx == SHN_ABS ||
473 S.st_value != SS.Value)
475 StringRef Name = check(S.getName(File.getStringTable()));
476 Symbol *Sym = Symtab->find(Name);
477 if (auto *Alias = dyn_cast_or_null<SharedSymbol>(Sym))
483 // When a symbol is copy relocated or we create a canonical plt entry, it is
484 // effectively a defined symbol. In the case of copy relocation the symbol is
485 // in .bss and in the case of a canonical plt entry it is in .plt. This function
486 // replaces the existing symbol with a Defined pointing to the appropriate
488 static void replaceWithDefined(Symbol &Sym, SectionBase *Sec, uint64_t Value,
491 replaceSymbol<Defined>(&Sym, Sym.File, Sym.getName(), Sym.Binding,
492 Sym.StOther, Sym.Type, Value, Size, Sec);
493 Sym.PltIndex = Old.PltIndex;
494 Sym.GotIndex = Old.GotIndex;
495 Sym.VerdefIndex = Old.VerdefIndex;
496 Sym.IsPreemptible = true;
497 Sym.ExportDynamic = true;
498 Sym.IsUsedInRegularObj = true;
502 // Reserve space in .bss or .bss.rel.ro for copy relocation.
504 // The copy relocation is pretty much a hack. If you use a copy relocation
505 // in your program, not only the symbol name but the symbol's size, RW/RO
506 // bit and alignment become part of the ABI. In addition to that, if the
507 // symbol has aliases, the aliases become part of the ABI. That's subtle,
508 // but if you violate that implicit ABI, that can cause very counter-
509 // intuitive consequences.
511 // So, what is the copy relocation? It's for linking non-position
512 // independent code to DSOs. In an ideal world, all references to data
513 // exported by DSOs should go indirectly through GOT. But if object files
514 // are compiled as non-PIC, all data references are direct. There is no
515 // way for the linker to transform the code to use GOT, as machine
516 // instructions are already set in stone in object files. This is where
517 // the copy relocation takes a role.
519 // A copy relocation instructs the dynamic linker to copy data from a DSO
520 // to a specified address (which is usually in .bss) at load-time. If the
521 // static linker (that's us) finds a direct data reference to a DSO
522 // symbol, it creates a copy relocation, so that the symbol can be
523 // resolved as if it were in .bss rather than in a DSO.
525 // As you can see in this function, we create a copy relocation for the
526 // dynamic linker, and the relocation contains not only symbol name but
527 // various other informtion about the symbol. So, such attributes become a
530 // Note for application developers: I can give you a piece of advice if
531 // you are writing a shared library. You probably should export only
532 // functions from your library. You shouldn't export variables.
534 // As an example what can happen when you export variables without knowing
535 // the semantics of copy relocations, assume that you have an exported
536 // variable of type T. It is an ABI-breaking change to add new members at
537 // end of T even though doing that doesn't change the layout of the
538 // existing members. That's because the space for the new members are not
539 // reserved in .bss unless you recompile the main program. That means they
540 // are likely to overlap with other data that happens to be laid out next
541 // to the variable in .bss. This kind of issue is sometimes very hard to
542 // debug. What's a solution? Instead of exporting a varaible V from a DSO,
543 // define an accessor getV().
544 template <class ELFT> static void addCopyRelSymbol(SharedSymbol &SS) {
545 // Copy relocation against zero-sized symbol doesn't make sense.
546 uint64_t SymSize = SS.getSize();
547 if (SymSize == 0 || SS.Alignment == 0)
548 fatal("cannot create a copy relocation for symbol " + toString(SS));
550 // See if this symbol is in a read-only segment. If so, preserve the symbol's
551 // memory protection by reserving space in the .bss.rel.ro section.
552 bool IsReadOnly = isReadOnly<ELFT>(SS);
553 BssSection *Sec = make<BssSection>(IsReadOnly ? ".bss.rel.ro" : ".bss",
554 SymSize, SS.Alignment);
556 InX::BssRelRo->getParent()->addSection(Sec);
558 InX::Bss->getParent()->addSection(Sec);
560 // Look through the DSO's dynamic symbol table for aliases and create a
561 // dynamic symbol for each one. This causes the copy relocation to correctly
562 // interpose any aliases.
563 for (SharedSymbol *Sym : getSymbolsAt<ELFT>(SS))
564 replaceWithDefined(*Sym, Sec, 0, Sym->Size);
566 InX::RelaDyn->addReloc(Target->CopyRel, Sec, 0, &SS);
569 // MIPS has an odd notion of "paired" relocations to calculate addends.
570 // For example, if a relocation is of R_MIPS_HI16, there must be a
571 // R_MIPS_LO16 relocation after that, and an addend is calculated using
572 // the two relocations.
573 template <class ELFT, class RelTy>
574 static int64_t computeMipsAddend(const RelTy &Rel, const RelTy *End,
575 InputSectionBase &Sec, RelExpr Expr,
577 if (Expr == R_MIPS_GOTREL && IsLocal)
578 return Sec.getFile<ELFT>()->MipsGp0;
580 // The ABI says that the paired relocation is used only for REL.
581 // See p. 4-17 at ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
585 RelType Type = Rel.getType(Config->IsMips64EL);
586 uint32_t PairTy = getMipsPairType(Type, IsLocal);
587 if (PairTy == R_MIPS_NONE)
590 const uint8_t *Buf = Sec.Data.data();
591 uint32_t SymIndex = Rel.getSymbol(Config->IsMips64EL);
593 // To make things worse, paired relocations might not be contiguous in
594 // the relocation table, so we need to do linear search. *sigh*
595 for (const RelTy *RI = &Rel; RI != End; ++RI)
596 if (RI->getType(Config->IsMips64EL) == PairTy &&
597 RI->getSymbol(Config->IsMips64EL) == SymIndex)
598 return Target->getImplicitAddend(Buf + RI->r_offset, PairTy);
600 warn("can't find matching " + toString(PairTy) + " relocation for " +
605 // Returns an addend of a given relocation. If it is RELA, an addend
606 // is in a relocation itself. If it is REL, we need to read it from an
608 template <class ELFT, class RelTy>
609 static int64_t computeAddend(const RelTy &Rel, const RelTy *End,
610 InputSectionBase &Sec, RelExpr Expr,
613 RelType Type = Rel.getType(Config->IsMips64EL);
616 Addend = getAddend<ELFT>(Rel);
618 const uint8_t *Buf = Sec.Data.data();
619 Addend = Target->getImplicitAddend(Buf + Rel.r_offset, Type);
622 if (Config->EMachine == EM_PPC64 && Config->Pic && Type == R_PPC64_TOC)
623 Addend += getPPC64TocBase();
624 if (Config->EMachine == EM_MIPS)
625 Addend += computeMipsAddend<ELFT>(Rel, End, Sec, Expr, IsLocal);
630 // Report an undefined symbol if necessary.
631 // Returns true if this function printed out an error message.
632 static bool maybeReportUndefined(Symbol &Sym, InputSectionBase &Sec,
634 if (Config->UnresolvedSymbols == UnresolvedPolicy::IgnoreAll)
637 if (Sym.isLocal() || !Sym.isUndefined() || Sym.isWeak())
641 Sym.computeBinding() != STB_LOCAL && Sym.Visibility == STV_DEFAULT;
642 if (Config->UnresolvedSymbols == UnresolvedPolicy::Ignore && CanBeExternal)
646 "undefined symbol: " + toString(Sym) + "\n>>> referenced by ";
648 std::string Src = Sec.getSrcMsg(Sym, Offset);
650 Msg += Src + "\n>>> ";
651 Msg += Sec.getObjMsg(Offset);
653 if ((Config->UnresolvedSymbols == UnresolvedPolicy::Warn && CanBeExternal) ||
654 Config->NoinhibitExec) {
663 // MIPS N32 ABI treats series of successive relocations with the same offset
664 // as a single relocation. The similar approach used by N64 ABI, but this ABI
665 // packs all relocations into the single relocation record. Here we emulate
666 // this for the N32 ABI. Iterate over relocation with the same offset and put
667 // theirs types into the single bit-set.
668 template <class RelTy> static RelType getMipsN32RelType(RelTy *&Rel, RelTy *End) {
670 uint64_t Offset = Rel->r_offset;
673 while (Rel != End && Rel->r_offset == Offset)
674 Type |= (Rel++)->getType(Config->IsMips64EL) << (8 * N++);
678 // .eh_frame sections are mergeable input sections, so their input
679 // offsets are not linearly mapped to output section. For each input
680 // offset, we need to find a section piece containing the offset and
681 // add the piece's base address to the input offset to compute the
682 // output offset. That isn't cheap.
684 // This class is to speed up the offset computation. When we process
685 // relocations, we access offsets in the monotonically increasing
686 // order. So we can optimize for that access pattern.
688 // For sections other than .eh_frame, this class doesn't do anything.
692 explicit OffsetGetter(InputSectionBase &Sec) {
693 if (auto *Eh = dyn_cast<EhInputSection>(&Sec))
697 // Translates offsets in input sections to offsets in output sections.
698 // Given offset must increase monotonically. We assume that Piece is
699 // sorted by InputOff.
700 uint64_t get(uint64_t Off) {
704 while (I != Pieces.size() && Pieces[I].InputOff + Pieces[I].Size <= Off)
706 if (I == Pieces.size())
709 // Pieces must be contiguous, so there must be no holes in between.
710 assert(Pieces[I].InputOff <= Off && "Relocation not in any piece");
712 // Offset -1 means that the piece is dead (i.e. garbage collected).
713 if (Pieces[I].OutputOff == -1)
715 return Pieces[I].OutputOff + Off - Pieces[I].InputOff;
719 ArrayRef<EhSectionPiece> Pieces;
724 static void addRelativeReloc(InputSectionBase *IS, uint64_t OffsetInSec,
725 Symbol *Sym, int64_t Addend, RelExpr Expr,
727 // Add a relative relocation. If RelrDyn section is enabled, and the
728 // relocation offset is guaranteed to be even, add the relocation to
729 // the RelrDyn section, otherwise add it to the RelaDyn section.
730 // RelrDyn sections don't support odd offsets. Also, RelrDyn sections
731 // don't store the addend values, so we must write it to the relocated
733 if (InX::RelrDyn && IS->Alignment >= 2 && OffsetInSec % 2 == 0) {
734 IS->Relocations.push_back({Expr, Type, OffsetInSec, Addend, Sym});
735 InX::RelrDyn->Relocs.push_back({IS, OffsetInSec});
738 InX::RelaDyn->addReloc(Target->RelativeRel, IS, OffsetInSec, Sym, Addend,
742 template <class ELFT, class GotPltSection>
743 static void addPltEntry(PltSection *Plt, GotPltSection *GotPlt,
744 RelocationBaseSection *Rel, RelType Type, Symbol &Sym) {
745 Plt->addEntry<ELFT>(Sym);
746 GotPlt->addEntry(Sym);
748 {Type, GotPlt, Sym.getGotPltOffset(), !Sym.IsPreemptible, &Sym, 0});
751 template <class ELFT> static void addGotEntry(Symbol &Sym) {
752 InX::Got->addEntry(Sym);
754 RelExpr Expr = Sym.isTls() ? R_TLS : R_ABS;
755 uint64_t Off = Sym.getGotOffset();
757 // If a GOT slot value can be calculated at link-time, which is now,
758 // we can just fill that out.
760 // (We don't actually write a value to a GOT slot right now, but we
761 // add a static relocation to a Relocations vector so that
762 // InputSection::relocate will do the work for us. We may be able
763 // to just write a value now, but it is a TODO.)
764 bool IsLinkTimeConstant =
765 !Sym.IsPreemptible && (!Config->Pic || isAbsolute(Sym));
766 if (IsLinkTimeConstant) {
767 InX::Got->Relocations.push_back({Expr, Target->GotRel, Off, 0, &Sym});
771 // Otherwise, we emit a dynamic relocation to .rel[a].dyn so that
772 // the GOT slot will be fixed at load-time.
773 if (!Sym.isTls() && !Sym.IsPreemptible && Config->Pic && !isAbsolute(Sym)) {
774 addRelativeReloc(InX::Got, Off, &Sym, 0, R_ABS, Target->GotRel);
777 InX::RelaDyn->addReloc(Sym.isTls() ? Target->TlsGotRel : Target->GotRel,
778 InX::Got, Off, &Sym, 0,
779 Sym.IsPreemptible ? R_ADDEND : R_ABS, Target->GotRel);
782 // Return true if we can define a symbol in the executable that
783 // contains the value/function of a symbol defined in a shared
785 static bool canDefineSymbolInExecutable(Symbol &Sym) {
786 // If the symbol has default visibility the symbol defined in the
787 // executable will preempt it.
788 // Note that we want the visibility of the shared symbol itself, not
789 // the visibility of the symbol in the output file we are producing. That is
790 // why we use Sym.StOther.
791 if ((Sym.StOther & 0x3) == STV_DEFAULT)
794 // If we are allowed to break address equality of functions, defining
795 // a plt entry will allow the program to call the function in the
796 // .so, but the .so and the executable will no agree on the address
797 // of the function. Similar logic for objects.
798 return ((Sym.isFunc() && Config->IgnoreFunctionAddressEquality) ||
799 (Sym.isObject() && Config->IgnoreDataAddressEquality));
802 // The reason we have to do this early scan is as follows
803 // * To mmap the output file, we need to know the size
804 // * For that, we need to know how many dynamic relocs we will have.
805 // It might be possible to avoid this by outputting the file with write:
806 // * Write the allocated output sections, computing addresses.
807 // * Apply relocations, recording which ones require a dynamic reloc.
808 // * Write the dynamic relocations.
809 // * Write the rest of the file.
810 // This would have some drawbacks. For example, we would only know if .rela.dyn
811 // is needed after applying relocations. If it is, it will go after rw and rx
812 // sections. Given that it is ro, we will need an extra PT_LOAD. This
813 // complicates things for the dynamic linker and means we would have to reserve
814 // space for the extra PT_LOAD even if we end up not using it.
815 template <class ELFT, class RelTy>
816 static void processRelocAux(InputSectionBase &Sec, RelExpr Expr, RelType Type,
817 uint64_t Offset, Symbol &Sym, const RelTy &Rel,
819 if (isStaticLinkTimeConstant(Expr, Type, Sym, Sec, Offset)) {
820 Sec.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
823 if (Sym.isGnuIFunc() && Config->ZIfuncnoplt) {
824 InX::RelaDyn->addReloc(Type, &Sec, Offset, &Sym, Addend, R_ADDEND, Type);
827 bool CanWrite = (Sec.Flags & SHF_WRITE) || !Config->ZText;
829 // R_GOT refers to a position in the got, even if the symbol is preemptible.
830 bool IsPreemptibleValue = Sym.IsPreemptible && Expr != R_GOT;
832 if (!IsPreemptibleValue) {
833 addRelativeReloc(&Sec, Offset, &Sym, Addend, Expr, Type);
835 } else if (RelType Rel = Target->getDynRel(Type)) {
836 InX::RelaDyn->addReloc(Rel, &Sec, Offset, &Sym, Addend, R_ADDEND, Type);
838 // MIPS ABI turns using of GOT and dynamic relocations inside out.
839 // While regular ABI uses dynamic relocations to fill up GOT entries
840 // MIPS ABI requires dynamic linker to fills up GOT entries using
841 // specially sorted dynamic symbol table. This affects even dynamic
842 // relocations against symbols which do not require GOT entries
843 // creation explicitly, i.e. do not have any GOT-relocations. So if
844 // a preemptible symbol has a dynamic relocation we anyway have
845 // to create a GOT entry for it.
846 // If a non-preemptible symbol has a dynamic relocation against it,
847 // dynamic linker takes it st_value, adds offset and writes down
848 // result of the dynamic relocation. In case of preemptible symbol
849 // dynamic linker performs symbol resolution, writes the symbol value
850 // to the GOT entry and reads the GOT entry when it needs to perform
851 // a dynamic relocation.
852 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf p.4-19
853 if (Config->EMachine == EM_MIPS)
854 InX::MipsGot->addEntry(*Sec.File, Sym, Addend, Expr);
859 // If the relocation is to a weak undef, and we are producing
860 // executable, give up on it and produce a non preemptible 0.
861 if (!Config->Shared && Sym.isUndefWeak()) {
862 Sec.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
866 if (!CanWrite && (Config->Pic && !isRelExpr(Expr))) {
868 "can't create dynamic relocation " + toString(Type) + " against " +
869 (Sym.getName().empty() ? "local symbol" : "symbol: " + toString(Sym)) +
870 " in readonly segment; recompile object files with -fPIC "
871 "or pass '-Wl,-z,notext' to allow text relocations in the output" +
872 getLocation(Sec, Sym, Offset));
876 // Copy relocations are only possible if we are creating an executable.
877 if (Config->Shared) {
878 errorOrWarn("relocation " + toString(Type) +
879 " cannot be used against symbol " + toString(Sym) +
880 "; recompile with -fPIC" + getLocation(Sec, Sym, Offset));
884 // If the symbol is undefined we already reported any relevant errors.
885 if (Sym.isUndefined())
888 if (!canDefineSymbolInExecutable(Sym)) {
889 error("cannot preempt symbol: " + toString(Sym) +
890 getLocation(Sec, Sym, Offset));
894 if (Sym.isObject()) {
895 // Produce a copy relocation.
896 if (auto *SS = dyn_cast<SharedSymbol>(&Sym)) {
897 if (!Config->ZCopyreloc)
898 error("unresolvable relocation " + toString(Type) +
899 " against symbol '" + toString(*SS) +
900 "'; recompile with -fPIC or remove '-z nocopyreloc'" +
901 getLocation(Sec, Sym, Offset));
902 addCopyRelSymbol<ELFT>(*SS);
904 Sec.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
909 // This handles a non PIC program call to function in a shared library. In
910 // an ideal world, we could just report an error saying the relocation can
911 // overflow at runtime. In the real world with glibc, crt1.o has a
912 // R_X86_64_PC32 pointing to libc.so.
914 // The general idea on how to handle such cases is to create a PLT entry and
915 // use that as the function value.
917 // For the static linking part, we just return a plt expr and everything
918 // else will use the PLT entry as the address.
920 // The remaining problem is making sure pointer equality still works. We
921 // need the help of the dynamic linker for that. We let it know that we have
922 // a direct reference to a so symbol by creating an undefined symbol with a
923 // non zero st_value. Seeing that, the dynamic linker resolves the symbol to
924 // the value of the symbol we created. This is true even for got entries, so
925 // pointer equality is maintained. To avoid an infinite loop, the only entry
926 // that points to the real function is a dedicated got entry used by the
927 // plt. That is identified by special relocation types (R_X86_64_JUMP_SLOT,
928 // R_386_JMP_SLOT, etc).
930 // For position independent executable on i386, the plt entry requires ebx
931 // to be set. This causes two problems:
932 // * If some code has a direct reference to a function, it was probably
933 // compiled without -fPIE/-fPIC and doesn't maintain ebx.
934 // * If a library definition gets preempted to the executable, it will have
935 // the wrong ebx value.
936 if (Config->Pie && Config->EMachine == EM_386)
937 errorOrWarn("symbol '" + toString(Sym) +
938 "' cannot be preempted; recompile with -fPIE" +
939 getLocation(Sec, Sym, Offset));
941 addPltEntry<ELFT>(InX::Plt, InX::GotPlt, InX::RelaPlt, Target->PltRel,
943 if (!Sym.isDefined())
944 replaceWithDefined(Sym, InX::Plt, Sym.getPltOffset(), 0);
945 Sym.NeedsPltAddr = true;
946 Sec.Relocations.push_back({Expr, Type, Offset, Addend, &Sym});
950 errorOrWarn("symbol '" + toString(Sym) + "' has no type" +
951 getLocation(Sec, Sym, Offset));
954 template <class ELFT, class RelTy>
955 static void scanReloc(InputSectionBase &Sec, OffsetGetter &GetOffset, RelTy *&I,
957 const RelTy &Rel = *I;
958 Symbol &Sym = Sec.getFile<ELFT>()->getRelocTargetSym(Rel);
961 // Deal with MIPS oddity.
962 if (Config->MipsN32Abi) {
963 Type = getMipsN32RelType(I, End);
965 Type = Rel.getType(Config->IsMips64EL);
969 // Get an offset in an output section this relocation is applied to.
970 uint64_t Offset = GetOffset.get(Rel.r_offset);
971 if (Offset == uint64_t(-1))
974 // Skip if the target symbol is an erroneous undefined symbol.
975 if (maybeReportUndefined(Sym, Sec, Rel.r_offset))
978 const uint8_t *RelocatedAddr = Sec.Data.begin() + Rel.r_offset;
979 RelExpr Expr = Target->getRelExpr(Type, Sym, RelocatedAddr);
981 // Ignore "hint" relocations because they are only markers for relaxation.
982 if (isRelExprOneOf<R_HINT, R_NONE>(Expr))
985 // Strenghten or relax relocations.
987 // GNU ifunc symbols must be accessed via PLT because their addresses
988 // are determined by runtime.
990 // On the other hand, if we know that a PLT entry will be resolved within
991 // the same ELF module, we can skip PLT access and directly jump to the
992 // destination function. For example, if we are linking a main exectuable,
993 // all dynamic symbols that can be resolved within the executable will
994 // actually be resolved that way at runtime, because the main exectuable
995 // is always at the beginning of a search list. We can leverage that fact.
996 if (Sym.isGnuIFunc() && !Config->ZIfuncnoplt)
998 else if (!Sym.IsPreemptible && Expr == R_GOT_PC && !isAbsoluteValue(Sym))
999 Expr = Target->adjustRelaxExpr(Type, RelocatedAddr, Expr);
1000 else if (!Sym.IsPreemptible)
1001 Expr = fromPlt(Expr);
1003 // This relocation does not require got entry, but it is relative to got and
1004 // needs it to be created. Here we request for that.
1005 if (isRelExprOneOf<R_GOTONLY_PC, R_GOTONLY_PC_FROM_END, R_GOTREL,
1006 R_GOTREL_FROM_END, R_PPC_TOC>(Expr))
1007 InX::Got->HasGotOffRel = true;
1010 int64_t Addend = computeAddend<ELFT>(Rel, End, Sec, Expr, Sym.isLocal());
1012 // Process some TLS relocations, including relaxing TLS relocations.
1013 // Note that this function does not handle all TLS relocations.
1014 if (unsigned Processed =
1015 handleTlsRelocation<ELFT>(Type, Sym, Sec, Offset, Addend, Expr)) {
1016 I += (Processed - 1);
1020 // If a relocation needs PLT, we create PLT and GOTPLT slots for the symbol.
1021 if (needsPlt(Expr) && !Sym.isInPlt()) {
1022 if (Sym.isGnuIFunc() && !Sym.IsPreemptible)
1023 addPltEntry<ELFT>(InX::Iplt, InX::IgotPlt, InX::RelaIplt,
1024 Target->IRelativeRel, Sym);
1026 addPltEntry<ELFT>(InX::Plt, InX::GotPlt, InX::RelaPlt, Target->PltRel,
1030 // Create a GOT slot if a relocation needs GOT.
1031 if (needsGot(Expr)) {
1032 if (Config->EMachine == EM_MIPS) {
1033 // MIPS ABI has special rules to process GOT entries and doesn't
1034 // require relocation entries for them. A special case is TLS
1035 // relocations. In that case dynamic loader applies dynamic
1036 // relocations to initialize TLS GOT entries.
1037 // See "Global Offset Table" in Chapter 5 in the following document
1038 // for detailed description:
1039 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
1040 InX::MipsGot->addEntry(*Sec.File, Sym, Addend, Expr);
1041 } else if (!Sym.isInGot()) {
1042 addGotEntry<ELFT>(Sym);
1046 processRelocAux<ELFT>(Sec, Expr, Type, Offset, Sym, Rel, Addend);
1049 template <class ELFT, class RelTy>
1050 static void scanRelocs(InputSectionBase &Sec, ArrayRef<RelTy> Rels) {
1051 OffsetGetter GetOffset(Sec);
1053 // Not all relocations end up in Sec.Relocations, but a lot do.
1054 Sec.Relocations.reserve(Rels.size());
1056 for (auto I = Rels.begin(), End = Rels.end(); I != End;)
1057 scanReloc<ELFT>(Sec, GetOffset, I, End);
1060 template <class ELFT> void elf::scanRelocations(InputSectionBase &S) {
1061 if (S.AreRelocsRela)
1062 scanRelocs<ELFT>(S, S.relas<ELFT>());
1064 scanRelocs<ELFT>(S, S.rels<ELFT>());
1067 // Thunk Implementation
1069 // Thunks (sometimes called stubs, veneers or branch islands) are small pieces
1070 // of code that the linker inserts inbetween a caller and a callee. The thunks
1071 // are added at link time rather than compile time as the decision on whether
1072 // a thunk is needed, such as the caller and callee being out of range, can only
1073 // be made at link time.
1075 // It is straightforward to tell given the current state of the program when a
1076 // thunk is needed for a particular call. The more difficult part is that
1077 // the thunk needs to be placed in the program such that the caller can reach
1078 // the thunk and the thunk can reach the callee; furthermore, adding thunks to
1079 // the program alters addresses, which can mean more thunks etc.
1081 // In lld we have a synthetic ThunkSection that can hold many Thunks.
1082 // The decision to have a ThunkSection act as a container means that we can
1083 // more easily handle the most common case of a single block of contiguous
1084 // Thunks by inserting just a single ThunkSection.
1086 // The implementation of Thunks in lld is split across these areas
1087 // Relocations.cpp : Framework for creating and placing thunks
1088 // Thunks.cpp : The code generated for each supported thunk
1089 // Target.cpp : Target specific hooks that the framework uses to decide when
1091 // Synthetic.cpp : Implementation of ThunkSection
1092 // Writer.cpp : Iteratively call framework until no more Thunks added
1094 // Thunk placement requirements:
1095 // Mips LA25 thunks. These must be placed immediately before the callee section
1096 // We can assume that the caller is in range of the Thunk. These are modelled
1097 // by Thunks that return the section they must precede with
1098 // getTargetInputSection().
1100 // ARM interworking and range extension thunks. These thunks must be placed
1101 // within range of the caller. All implemented ARM thunks can always reach the
1102 // callee as they use an indirect jump via a register that has no range
1105 // Thunk placement algorithm:
1106 // For Mips LA25 ThunkSections; the placement is explicit, it has to be before
1107 // getTargetInputSection().
1109 // For thunks that must be placed within range of the caller there are many
1110 // possible choices given that the maximum range from the caller is usually
1111 // much larger than the average InputSection size. Desirable properties include:
1112 // - Maximize reuse of thunks by multiple callers
1113 // - Minimize number of ThunkSections to simplify insertion
1114 // - Handle impact of already added Thunks on addresses
1115 // - Simple to understand and implement
1117 // In lld for the first pass, we pre-create one or more ThunkSections per
1118 // InputSectionDescription at Target specific intervals. A ThunkSection is
1119 // placed so that the estimated end of the ThunkSection is within range of the
1120 // start of the InputSectionDescription or the previous ThunkSection. For
1122 // InputSectionDescription
1132 // The intention is that we can add a Thunk to a ThunkSection that is well
1133 // spaced enough to service a number of callers without having to do a lot
1134 // of work. An important principle is that it is not an error if a Thunk cannot
1135 // be placed in a pre-created ThunkSection; when this happens we create a new
1136 // ThunkSection placed next to the caller. This allows us to handle the vast
1137 // majority of thunks simply, but also handle rare cases where the branch range
1138 // is smaller than the target specific spacing.
1140 // The algorithm is expected to create all the thunks that are needed in a
1141 // single pass, with a small number of programs needing a second pass due to
1142 // the insertion of thunks in the first pass increasing the offset between
1143 // callers and callees that were only just in range.
1145 // A consequence of allowing new ThunkSections to be created outside of the
1146 // pre-created ThunkSections is that in rare cases calls to Thunks that were in
1147 // range in pass K, are out of range in some pass > K due to the insertion of
1148 // more Thunks in between the caller and callee. When this happens we retarget
1149 // the relocation back to the original target and create another Thunk.
1151 // Remove ThunkSections that are empty, this should only be the initial set
1152 // precreated on pass 0.
1154 // Insert the Thunks for OutputSection OS into their designated place
1155 // in the Sections vector, and recalculate the InputSection output section
1157 // This may invalidate any output section offsets stored outside of InputSection
1158 void ThunkCreator::mergeThunks(ArrayRef<OutputSection *> OutputSections) {
1159 forEachInputSectionDescription(
1160 OutputSections, [&](OutputSection *OS, InputSectionDescription *ISD) {
1161 if (ISD->ThunkSections.empty())
1164 // Remove any zero sized precreated Thunks.
1165 llvm::erase_if(ISD->ThunkSections,
1166 [](const std::pair<ThunkSection *, uint32_t> &TS) {
1167 return TS.first->getSize() == 0;
1169 // ISD->ThunkSections contains all created ThunkSections, including
1170 // those inserted in previous passes. Extract the Thunks created this
1171 // pass and order them in ascending OutSecOff.
1172 std::vector<ThunkSection *> NewThunks;
1173 for (const std::pair<ThunkSection *, uint32_t> TS : ISD->ThunkSections)
1174 if (TS.second == Pass)
1175 NewThunks.push_back(TS.first);
1176 std::stable_sort(NewThunks.begin(), NewThunks.end(),
1177 [](const ThunkSection *A, const ThunkSection *B) {
1178 return A->OutSecOff < B->OutSecOff;
1181 // Merge sorted vectors of Thunks and InputSections by OutSecOff
1182 std::vector<InputSection *> Tmp;
1183 Tmp.reserve(ISD->Sections.size() + NewThunks.size());
1184 auto MergeCmp = [](const InputSection *A, const InputSection *B) {
1185 // std::merge requires a strict weak ordering.
1186 if (A->OutSecOff < B->OutSecOff)
1188 if (A->OutSecOff == B->OutSecOff) {
1189 auto *TA = dyn_cast<ThunkSection>(A);
1190 auto *TB = dyn_cast<ThunkSection>(B);
1191 // Check if Thunk is immediately before any specific Target
1192 // InputSection for example Mips LA25 Thunks.
1193 if (TA && TA->getTargetInputSection() == B)
1195 if (TA && !TB && !TA->getTargetInputSection())
1196 // Place Thunk Sections without specific targets before
1197 // non-Thunk Sections.
1202 std::merge(ISD->Sections.begin(), ISD->Sections.end(),
1203 NewThunks.begin(), NewThunks.end(), std::back_inserter(Tmp),
1205 ISD->Sections = std::move(Tmp);
1209 // Find or create a ThunkSection within the InputSectionDescription (ISD) that
1210 // is in range of Src. An ISD maps to a range of InputSections described by a
1211 // linker script section pattern such as { .text .text.* }.
1212 ThunkSection *ThunkCreator::getISDThunkSec(OutputSection *OS, InputSection *IS,
1213 InputSectionDescription *ISD,
1214 uint32_t Type, uint64_t Src) {
1215 for (std::pair<ThunkSection *, uint32_t> TP : ISD->ThunkSections) {
1216 ThunkSection *TS = TP.first;
1217 uint64_t TSBase = OS->Addr + TS->OutSecOff;
1218 uint64_t TSLimit = TSBase + TS->getSize();
1219 if (Target->inBranchRange(Type, Src, (Src > TSLimit) ? TSBase : TSLimit))
1223 // No suitable ThunkSection exists. This can happen when there is a branch
1224 // with lower range than the ThunkSection spacing or when there are too
1225 // many Thunks. Create a new ThunkSection as close to the InputSection as
1226 // possible. Error if InputSection is so large we cannot place ThunkSection
1227 // anywhere in Range.
1228 uint64_t ThunkSecOff = IS->OutSecOff;
1229 if (!Target->inBranchRange(Type, Src, OS->Addr + ThunkSecOff)) {
1230 ThunkSecOff = IS->OutSecOff + IS->getSize();
1231 if (!Target->inBranchRange(Type, Src, OS->Addr + ThunkSecOff))
1232 fatal("InputSection too large for range extension thunk " +
1233 IS->getObjMsg(Src - (OS->Addr + IS->OutSecOff)));
1235 return addThunkSection(OS, ISD, ThunkSecOff);
1238 // Add a Thunk that needs to be placed in a ThunkSection that immediately
1239 // precedes its Target.
1240 ThunkSection *ThunkCreator::getISThunkSec(InputSection *IS) {
1241 ThunkSection *TS = ThunkedSections.lookup(IS);
1245 // Find InputSectionRange within Target Output Section (TOS) that the
1246 // InputSection (IS) that we need to precede is in.
1247 OutputSection *TOS = IS->getParent();
1248 for (BaseCommand *BC : TOS->SectionCommands)
1249 if (auto *ISD = dyn_cast<InputSectionDescription>(BC)) {
1250 if (ISD->Sections.empty())
1252 InputSection *first = ISD->Sections.front();
1253 InputSection *last = ISD->Sections.back();
1254 if (IS->OutSecOff >= first->OutSecOff &&
1255 IS->OutSecOff <= last->OutSecOff) {
1256 TS = addThunkSection(TOS, ISD, IS->OutSecOff);
1257 ThunkedSections[IS] = TS;
1264 // Create one or more ThunkSections per OS that can be used to place Thunks.
1265 // We attempt to place the ThunkSections using the following desirable
1267 // - Within range of the maximum number of callers
1268 // - Minimise the number of ThunkSections
1270 // We follow a simple but conservative heuristic to place ThunkSections at
1271 // offsets that are multiples of a Target specific branch range.
1272 // For an InputSectionDescription that is smaller than the range, a single
1273 // ThunkSection at the end of the range will do.
1275 // For an InputSectionDescription that is more than twice the size of the range,
1276 // we place the last ThunkSection at range bytes from the end of the
1277 // InputSectionDescription in order to increase the likelihood that the
1278 // distance from a thunk to its target will be sufficiently small to
1279 // allow for the creation of a short thunk.
1280 void ThunkCreator::createInitialThunkSections(
1281 ArrayRef<OutputSection *> OutputSections) {
1282 forEachInputSectionDescription(
1283 OutputSections, [&](OutputSection *OS, InputSectionDescription *ISD) {
1284 if (ISD->Sections.empty())
1286 uint32_t ISDBegin = ISD->Sections.front()->OutSecOff;
1288 ISD->Sections.back()->OutSecOff + ISD->Sections.back()->getSize();
1289 uint32_t LastThunkLowerBound = -1;
1290 if (ISDEnd - ISDBegin > Target->ThunkSectionSpacing * 2)
1291 LastThunkLowerBound = ISDEnd - Target->ThunkSectionSpacing;
1294 uint32_t PrevISLimit = ISDBegin;
1295 uint32_t ThunkUpperBound = ISDBegin + Target->ThunkSectionSpacing;
1297 for (const InputSection *IS : ISD->Sections) {
1298 ISLimit = IS->OutSecOff + IS->getSize();
1299 if (ISLimit > ThunkUpperBound) {
1300 addThunkSection(OS, ISD, PrevISLimit);
1301 ThunkUpperBound = PrevISLimit + Target->ThunkSectionSpacing;
1303 if (ISLimit > LastThunkLowerBound)
1305 PrevISLimit = ISLimit;
1307 addThunkSection(OS, ISD, ISLimit);
1311 ThunkSection *ThunkCreator::addThunkSection(OutputSection *OS,
1312 InputSectionDescription *ISD,
1314 auto *TS = make<ThunkSection>(OS, Off);
1315 ISD->ThunkSections.push_back(std::make_pair(TS, Pass));
1319 std::pair<Thunk *, bool> ThunkCreator::getThunk(Symbol &Sym, RelType Type,
1321 std::vector<Thunk *> *ThunkVec = nullptr;
1322 // We use (section, offset) pair to find the thunk position if possible so
1323 // that we create only one thunk for aliased symbols or ICFed sections.
1324 if (auto *D = dyn_cast<Defined>(&Sym))
1325 if (!D->isInPlt() && D->Section)
1326 ThunkVec = &ThunkedSymbolsBySection[{D->Section->Repl, D->Value}];
1328 ThunkVec = &ThunkedSymbols[&Sym];
1329 // Check existing Thunks for Sym to see if they can be reused
1330 for (Thunk *ET : *ThunkVec)
1331 if (ET->isCompatibleWith(Type) &&
1332 Target->inBranchRange(Type, Src, ET->getThunkTargetSym()->getVA()))
1333 return std::make_pair(ET, false);
1334 // No existing compatible Thunk in range, create a new one
1335 Thunk *T = addThunk(Type, Sym);
1336 ThunkVec->push_back(T);
1337 return std::make_pair(T, true);
1340 // Call Fn on every executable InputSection accessed via the linker script
1341 // InputSectionDescription::Sections.
1342 void ThunkCreator::forEachInputSectionDescription(
1343 ArrayRef<OutputSection *> OutputSections,
1344 llvm::function_ref<void(OutputSection *, InputSectionDescription *)> Fn) {
1345 for (OutputSection *OS : OutputSections) {
1346 if (!(OS->Flags & SHF_ALLOC) || !(OS->Flags & SHF_EXECINSTR))
1348 for (BaseCommand *BC : OS->SectionCommands)
1349 if (auto *ISD = dyn_cast<InputSectionDescription>(BC))
1354 // Return true if the relocation target is an in range Thunk.
1355 // Return false if the relocation is not to a Thunk. If the relocation target
1356 // was originally to a Thunk, but is no longer in range we revert the
1357 // relocation back to its original non-Thunk target.
1358 bool ThunkCreator::normalizeExistingThunk(Relocation &Rel, uint64_t Src) {
1359 if (Thunk *ET = Thunks.lookup(Rel.Sym)) {
1360 if (Target->inBranchRange(Rel.Type, Src, Rel.Sym->getVA()))
1362 Rel.Sym = &ET->Destination;
1363 if (Rel.Sym->isInPlt())
1364 Rel.Expr = toPlt(Rel.Expr);
1369 // Process all relocations from the InputSections that have been assigned
1370 // to InputSectionDescriptions and redirect through Thunks if needed. The
1371 // function should be called iteratively until it returns false.
1374 // All InputSections that may need a Thunk are reachable from
1375 // OutputSectionCommands.
1377 // All OutputSections have an address and all InputSections have an offset
1378 // within the OutputSection.
1380 // The offsets between caller (relocation place) and callee
1381 // (relocation target) will not be modified outside of createThunks().
1384 // If return value is true then ThunkSections have been inserted into
1385 // OutputSections. All relocations that needed a Thunk based on the information
1386 // available to createThunks() on entry have been redirected to a Thunk. Note
1387 // that adding Thunks changes offsets between caller and callee so more Thunks
1390 // If return value is false then no more Thunks are needed, and createThunks has
1391 // made no changes. If the target requires range extension thunks, currently
1392 // ARM, then any future change in offset between caller and callee risks a
1393 // relocation out of range error.
1394 bool ThunkCreator::createThunks(ArrayRef<OutputSection *> OutputSections) {
1395 bool AddressesChanged = false;
1396 if (Pass == 0 && Target->ThunkSectionSpacing)
1397 createInitialThunkSections(OutputSections);
1398 else if (Pass == 10)
1399 // With Thunk Size much smaller than branch range we expect to
1400 // converge quickly; if we get to 10 something has gone wrong.
1401 fatal("thunk creation not converged");
1403 // Create all the Thunks and insert them into synthetic ThunkSections. The
1404 // ThunkSections are later inserted back into InputSectionDescriptions.
1405 // We separate the creation of ThunkSections from the insertion of the
1406 // ThunkSections as ThunkSections are not always inserted into the same
1407 // InputSectionDescription as the caller.
1408 forEachInputSectionDescription(
1409 OutputSections, [&](OutputSection *OS, InputSectionDescription *ISD) {
1410 for (InputSection *IS : ISD->Sections)
1411 for (Relocation &Rel : IS->Relocations) {
1412 uint64_t Src = IS->getVA(Rel.Offset);
1414 // If we are a relocation to an existing Thunk, check if it is
1415 // still in range. If not then Rel will be altered to point to its
1416 // original target so another Thunk can be generated.
1417 if (Pass > 0 && normalizeExistingThunk(Rel, Src))
1420 if (!Target->needsThunk(Rel.Expr, Rel.Type, IS->File, Src,
1425 std::tie(T, IsNew) = getThunk(*Rel.Sym, Rel.Type, Src);
1427 // Find or create a ThunkSection for the new Thunk
1429 if (auto *TIS = T->getTargetInputSection())
1430 TS = getISThunkSec(TIS);
1432 TS = getISDThunkSec(OS, IS, ISD, Rel.Type, Src);
1434 Thunks[T->getThunkTargetSym()] = T;
1436 // Redirect relocation to Thunk, we never go via the PLT to a Thunk
1437 Rel.Sym = T->getThunkTargetSym();
1438 Rel.Expr = fromPlt(Rel.Expr);
1440 for (auto &P : ISD->ThunkSections)
1441 AddressesChanged |= P.first->assignOffsets();
1443 for (auto &P : ThunkedSections)
1444 AddressesChanged |= P.second->assignOffsets();
1446 // Merge all created synthetic ThunkSections back into OutputSection
1447 mergeThunks(OutputSections);
1449 return AddressesChanged;
1452 template void elf::scanRelocations<ELF32LE>(InputSectionBase &);
1453 template void elf::scanRelocations<ELF32BE>(InputSectionBase &);
1454 template void elf::scanRelocations<ELF64LE>(InputSectionBase &);
1455 template void elf::scanRelocations<ELF64BE>(InputSectionBase &);