1//===- InputSection.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 "InputSection.h"
10#include "Config.h"
11#include "InputFiles.h"
12#include "OutputSections.h"
13#include "Relocations.h"
14#include "SymbolTable.h"
15#include "Symbols.h"
16#include "SyntheticSections.h"
17#include "Target.h"
18#include "lld/Common/DWARF.h"
19#include "llvm/Support/Compiler.h"
20#include "llvm/Support/Compression.h"
21#include "llvm/Support/Endian.h"
22#include "llvm/Support/xxhash.h"
23#include <algorithm>
24#include <optional>
25#include <vector>
26
27using namespace llvm;
28using namespace llvm::ELF;
29using namespace llvm::object;
30using namespace llvm::support;
31using namespace llvm::support::endian;
32using namespace llvm::sys;
33using namespace lld;
34using namespace lld::elf;
35
36// Returns a string to construct an error message.
37std::string elf::toStr(Ctx &ctx, const InputSectionBase *sec) {
38 return (toStr(ctx, f: sec->file) + ":(" + sec->name + ")").str();
39}
40
41const ELFSyncStream &elf::operator<<(const ELFSyncStream &s,
42 const InputSectionBase *sec) {
43 return s << toStr(ctx&: s.ctx, sec);
44}
45
46template <class ELFT>
47static ArrayRef<uint8_t> getSectionContents(ObjFile<ELFT> &file,
48 const typename ELFT::Shdr &hdr) {
49 if (hdr.sh_type == SHT_NOBITS)
50 return ArrayRef<uint8_t>(nullptr, hdr.sh_size);
51 return check(file.getObj().getSectionContents(hdr));
52}
53
54InputSectionBase::InputSectionBase(InputFile *file, StringRef name,
55 uint32_t type, uint64_t flags, uint32_t link,
56 uint32_t info, uint32_t addralign,
57 uint32_t entsize, ArrayRef<uint8_t> data,
58 Kind sectionKind)
59 : SectionBase(sectionKind, file, name, type, flags, link, info, addralign,
60 entsize),
61 bss(0), decodedCrel(0), keepUnique(0), nopFiller(0),
62 content_(data.data()), size(data.size()) {
63 // In order to reduce memory allocation, we assume that mergeable
64 // sections are smaller than 4 GiB, which is not an unreasonable
65 // assumption as of 2017.
66 if (sectionKind == SectionBase::Merge && content().size() > UINT32_MAX)
67 ErrAlways(ctx&: getCtx()) << this << ": section too large";
68
69 // The ELF spec states that a value of 0 means the section has
70 // no alignment constraints.
71 uint32_t v = std::max<uint32_t>(a: addralign, b: 1);
72 if (!isPowerOf2_64(Value: v)) {
73 Err(ctx&: getCtx()) << this << ": sh_addralign is not a power of 2";
74 v = 1;
75 }
76 this->addralign = v;
77
78 // If SHF_COMPRESSED is set, parse the header. The legacy .zdebug format is no
79 // longer supported.
80 if (flags & SHF_COMPRESSED) {
81 Ctx &ctx = file->ctx;
82 invokeELFT(parseCompressedHeader, ctx);
83 }
84}
85
86// SHF_INFO_LINK and SHF_GROUP are normally resolved and not copied to the
87// output section. However, for relocatable linking without
88// --force-group-allocation, the SHF_GROUP flag and section groups are retained.
89static uint64_t getFlags(Ctx &ctx, uint64_t flags) {
90 flags &= ~(uint64_t)SHF_INFO_LINK;
91 if (ctx.arg.resolveGroups)
92 flags &= ~(uint64_t)SHF_GROUP;
93 return flags;
94}
95
96template <class ELFT>
97InputSectionBase::InputSectionBase(ObjFile<ELFT> &file,
98 const typename ELFT::Shdr &hdr,
99 StringRef name, Kind sectionKind)
100 : InputSectionBase(&file, name, hdr.sh_type,
101 getFlags(file.ctx, hdr.sh_flags), hdr.sh_link,
102 hdr.sh_info, hdr.sh_addralign, hdr.sh_entsize,
103 getSectionContents(file, hdr), sectionKind) {
104 // We reject object files having insanely large alignments even though
105 // they are allowed by the spec. I think 4GB is a reasonable limitation.
106 // We might want to relax this in the future.
107 if (hdr.sh_addralign > UINT32_MAX) {
108 Err(ctx&: getCtx()) << &file << ": section sh_addralign is too large";
109 addralign = 1;
110 }
111}
112
113size_t InputSectionBase::getSize() const {
114 if (auto *s = dyn_cast<SyntheticSection>(Val: this))
115 return s->getSize();
116 return size - bytesDropped;
117}
118
119template <class ELFT>
120static void decompressAux(Ctx &ctx, const InputSectionBase &sec, uint8_t *out,
121 size_t size) {
122 auto *hdr = reinterpret_cast<const typename ELFT::Chdr *>(sec.content_);
123 auto compressed = ArrayRef<uint8_t>(sec.content_, sec.compressedSize)
124 .slice(N: sizeof(typename ELFT::Chdr));
125 if (Error e = hdr->ch_type == ELFCOMPRESS_ZLIB
126 ? compression::zlib::decompress(Input: compressed, Output: out, UncompressedSize&: size)
127 : compression::zstd::decompress(Input: compressed, Output: out, UncompressedSize&: size))
128 Err(ctx) << &sec << ": decompress failed: " << std::move(e);
129}
130
131void InputSectionBase::decompress() const {
132 Ctx &ctx = getCtx();
133 uint8_t *buf = makeThreadLocalN<uint8_t>(n: size);
134 invokeELFT(decompressAux, ctx, *this, buf, size);
135 content_ = buf;
136 compressed = false;
137}
138
139template <class ELFT>
140RelsOrRelas<ELFT> InputSectionBase::relsOrRelas(bool supportsCrel) const {
141 if (relSecIdx == 0)
142 return {};
143 RelsOrRelas<ELFT> ret;
144 auto *f = cast<ObjFile<ELFT>>(file);
145 typename ELFT::Shdr shdr = f->template getELFShdrs<ELFT>()[relSecIdx];
146 if (shdr.sh_type == SHT_CREL) {
147 // Return an iterator if supported by caller.
148 if (supportsCrel) {
149 ret.crels = Relocs<typename ELFT::Crel>(
150 (const uint8_t *)f->mb.getBufferStart() + shdr.sh_offset);
151 return ret;
152 }
153 InputSectionBase *const &relSec = f->getSections()[relSecIdx];
154 // Otherwise, allocate a buffer to hold the decoded RELA relocations. When
155 // called for the first time, relSec is null (without --emit-relocs) or an
156 // InputSection with false decodedCrel.
157 if (!relSec || !cast<InputSection>(Val: relSec)->decodedCrel) {
158 auto *sec = makeThreadLocal<InputSection>(*f, shdr, name);
159 f->cacheDecodedCrel(relSecIdx, sec);
160 sec->type = SHT_RELA;
161 sec->decodedCrel = true;
162
163 RelocsCrel<ELFT::Is64Bits> entries(sec->content_);
164 sec->size = entries.size() * sizeof(typename ELFT::Rela);
165 auto *relas = makeThreadLocalN<typename ELFT::Rela>(entries.size());
166 sec->content_ = reinterpret_cast<uint8_t *>(relas);
167 for (auto [i, r] : llvm::enumerate(entries)) {
168 relas[i].r_offset = r.r_offset;
169 relas[i].setSymbolAndType(r.r_symidx, r.r_type, false);
170 relas[i].r_addend = r.r_addend;
171 }
172 }
173 ret.relas = {ArrayRef(
174 reinterpret_cast<const typename ELFT::Rela *>(relSec->content_),
175 relSec->size / sizeof(typename ELFT::Rela))};
176 return ret;
177 }
178
179 const void *content = f->mb.getBufferStart() + shdr.sh_offset;
180 size_t size = shdr.sh_size;
181 if (shdr.sh_type == SHT_REL) {
182 ret.rels = {ArrayRef(reinterpret_cast<const typename ELFT::Rel *>(content),
183 size / sizeof(typename ELFT::Rel))};
184 } else {
185 assert(shdr.sh_type == SHT_RELA);
186 ret.relas = {
187 ArrayRef(reinterpret_cast<const typename ELFT::Rela *>(content),
188 size / sizeof(typename ELFT::Rela))};
189 }
190 return ret;
191}
192
193Ctx &SectionBase::getCtx() const { return file->ctx; }
194
195uint64_t SectionBase::getOffset(uint64_t offset) const {
196 switch (kind()) {
197 case Output: {
198 auto *os = cast<OutputSection>(Val: this);
199 // For output sections we treat offset -1 as the end of the section.
200 return offset == uint64_t(-1) ? os->size : offset;
201 }
202 case Class:
203 llvm_unreachable("section classes do not have offsets");
204 case Regular:
205 case Synthetic:
206 case Spill:
207 return cast<InputSection>(Val: this)->outSecOff + offset;
208 case EHFrame: {
209 // Two code paths may reach here. First, clang_rt.crtbegin.o and GCC
210 // crtbeginT.o may reference the start of an empty .eh_frame to identify the
211 // start of the output .eh_frame. Just return offset.
212 //
213 // Second, InputSection::copyRelocations on .eh_frame. Some pieces may be
214 // discarded due to GC/ICF. We should compute the output section offset.
215 const EhInputSection *es = cast<EhInputSection>(Val: this);
216 if (!es->content().empty())
217 if (InputSection *isec = es->getParent())
218 return isec->outSecOff + es->getParentOffset(offset);
219 return offset;
220 }
221 case Merge:
222 const MergeInputSection *ms = cast<MergeInputSection>(Val: this);
223 if (InputSection *isec = ms->getParent())
224 return isec->outSecOff + ms->getParentOffset(offset);
225 return ms->getParentOffset(offset);
226 }
227 llvm_unreachable("invalid section kind");
228}
229
230uint64_t SectionBase::getVA(uint64_t offset) const {
231 const OutputSection *out = getOutputSection();
232 return (out ? out->addr : 0) + getOffset(offset);
233}
234
235OutputSection *SectionBase::getOutputSection() {
236 InputSection *sec;
237 if (auto *isec = dyn_cast<InputSection>(Val: this))
238 sec = isec;
239 else if (auto *ms = dyn_cast<MergeInputSection>(Val: this))
240 sec = ms->getParent();
241 else if (auto *eh = dyn_cast<EhInputSection>(Val: this))
242 sec = eh->getParent();
243 else
244 return cast<OutputSection>(Val: this);
245 return sec ? sec->getParent() : nullptr;
246}
247
248// When a section is compressed, `rawData` consists with a header followed
249// by zlib-compressed data. This function parses a header to initialize
250// `uncompressedSize` member and remove the header from `rawData`.
251template <typename ELFT>
252void InputSectionBase::parseCompressedHeader(Ctx &ctx) {
253 flags &= ~(uint64_t)SHF_COMPRESSED;
254
255 // New-style header
256 if (content().size() < sizeof(typename ELFT::Chdr)) {
257 ErrAlways(ctx) << this << ": corrupted compressed section";
258 return;
259 }
260
261 auto *hdr = reinterpret_cast<const typename ELFT::Chdr *>(content().data());
262 if (hdr->ch_type == ELFCOMPRESS_ZLIB) {
263 if (!compression::zlib::isAvailable())
264 ErrAlways(ctx) << this
265 << " is compressed with ELFCOMPRESS_ZLIB, but lld is "
266 "not built with zlib support";
267 } else if (hdr->ch_type == ELFCOMPRESS_ZSTD) {
268 if (!compression::zstd::isAvailable())
269 ErrAlways(ctx) << this
270 << " is compressed with ELFCOMPRESS_ZSTD, but lld is "
271 "not built with zstd support";
272 } else {
273 ErrAlways(ctx) << this << ": unsupported compression type ("
274 << uint32_t(hdr->ch_type) << ")";
275 return;
276 }
277
278 compressed = true;
279 compressedSize = size;
280 size = hdr->ch_size;
281 addralign = std::max<uint32_t>(hdr->ch_addralign, 1);
282}
283
284InputSection *InputSectionBase::getLinkOrderDep() const {
285 assert(flags & SHF_LINK_ORDER);
286 if (!link)
287 return nullptr;
288 return cast<InputSection>(Val: file->getSections()[link]);
289}
290
291// Find a symbol that encloses a given location.
292Defined *InputSectionBase::getEnclosingSymbol(uint64_t offset,
293 uint8_t type) const {
294 if (file->isInternal())
295 return nullptr;
296 for (Symbol *b : file->getSymbols())
297 if (Defined *d = dyn_cast<Defined>(Val: b))
298 if (d->section == this && d->value <= offset &&
299 offset < d->value + d->size && (type == 0 || type == d->type))
300 return d;
301 return nullptr;
302}
303
304// Returns an object file location string. Used to construct an error message.
305std::string InputSectionBase::getLocation(uint64_t offset) const {
306 std::string secAndOffset =
307 (name + "+0x" + Twine::utohexstr(Val: offset) + ")").str();
308
309 std::string filename = toStr(getCtx(), f: file);
310 if (Defined *d = getEnclosingFunction(offset))
311 return filename + ":(function " + toStr(getCtx(), *d) + ": " + secAndOffset;
312
313 return filename + ":(" + secAndOffset;
314}
315
316static void printFileLine(const ELFSyncStream &s, StringRef path,
317 unsigned line) {
318 StringRef filename = path::filename(path);
319 s << filename << ':' << line;
320 if (filename != path)
321 s << " (" << path << ':' << line << ')';
322}
323
324// Print an error message that looks like this:
325//
326// foo.c:42 (/home/alice/possibly/very/long/path/foo.c:42)
327const ELFSyncStream &elf::operator<<(const ELFSyncStream &s,
328 InputSectionBase::SrcMsg &&msg) {
329 auto &sec = msg.sec;
330 if (sec.file->kind() != InputFile::ObjKind)
331 return s;
332 auto &file = cast<ELFFileBase>(Val&: *sec.file);
333
334 // First, look up the DWARF line table.
335 ArrayRef<InputSectionBase *> sections = file.getSections();
336 auto it = llvm::find(Range&: sections, Val: &sec);
337 uint64_t sectionIndex = it != sections.end()
338 ? it - sections.begin()
339 : object::SectionedAddress::UndefSection;
340 DWARFCache *dwarf = file.getDwarf();
341 if (auto info = dwarf->getDILineInfo(offset: msg.offset, sectionIndex))
342 printFileLine(s, path: info->FileName, line: info->Line);
343 else if (auto fileLine = dwarf->getVariableLoc(name: msg.sym.getName()))
344 // If it failed, look up again as a variable.
345 printFileLine(s, path: fileLine->first, line: fileLine->second);
346 else
347 // File.sourceFile contains STT_FILE symbol, and that is a last resort.
348 s << file.sourceFile;
349 return s;
350}
351
352// Returns a filename string along with an optional section name. This
353// function is intended to be used for constructing an error
354// message. The returned message looks like this:
355//
356// path/to/foo.o:(function bar)
357//
358// or
359//
360// path/to/foo.o:(function bar) in archive path/to/bar.a
361const ELFSyncStream &elf::operator<<(const ELFSyncStream &s,
362 InputSectionBase::ObjMsg &&msg) {
363 auto *sec = msg.sec;
364 s << sec->file->getName() << ":(";
365
366 // Find a symbol that encloses a given location. getObjMsg may be called
367 // before ObjFile::initSectionsAndLocalSyms where local symbols are
368 // initialized.
369 if (Defined *d = sec->getEnclosingSymbol(offset: msg.offset))
370 s << d;
371 else
372 s << sec->name << "+0x" << Twine::utohexstr(Val: msg.offset);
373 s << ')';
374 if (!sec->file->archiveName.empty())
375 s << (" in archive " + sec->file->archiveName).str();
376 return s;
377}
378
379PotentialSpillSection::PotentialSpillSection(const InputSectionBase &source,
380 InputSectionDescription &isd)
381 : InputSection(source.file, source.name, source.type, source.flags,
382 source.addralign, source.addralign, {}, SectionBase::Spill),
383 isd(&isd) {}
384
385InputSection InputSection::discarded(nullptr, "", 0, 0, 0, 0,
386 ArrayRef<uint8_t>());
387
388InputSection::InputSection(InputFile *f, StringRef name, uint32_t type,
389 uint64_t flags, uint32_t addralign, uint32_t entsize,
390 ArrayRef<uint8_t> data, Kind k)
391 : InputSectionBase(f, name, type, flags,
392 /*link=*/0, /*info=*/0, addralign, /*entsize=*/entsize,
393 data, k) {
394 assert(f || this == &InputSection::discarded);
395}
396
397template <class ELFT>
398InputSection::InputSection(ObjFile<ELFT> &f, const typename ELFT::Shdr &header,
399 StringRef name)
400 : InputSectionBase(f, header, name, InputSectionBase::Regular) {}
401
402// Copy SHT_GROUP section contents. Used only for the -r option.
403template <class ELFT> void InputSection::copyShtGroup(uint8_t *buf) {
404 // ELFT::Word is the 32-bit integral type in the target endianness.
405 using u32 = typename ELFT::Word;
406 ArrayRef<u32> from = getDataAs<u32>();
407 auto *to = reinterpret_cast<u32 *>(buf);
408
409 // The first entry is not a section number but a flag.
410 *to++ = from[0];
411
412 // Adjust section numbers because section numbers in an input object files are
413 // different in the output. We also need to handle combined or discarded
414 // members.
415 ArrayRef<InputSectionBase *> sections = file->getSections();
416 DenseSet<uint32_t> seen;
417 for (uint32_t idx : from.slice(1)) {
418 OutputSection *osec = sections[idx]->getOutputSection();
419 if (osec && seen.insert(V: osec->sectionIndex).second)
420 *to++ = osec->sectionIndex;
421 }
422}
423
424InputSectionBase *InputSection::getRelocatedSection() const {
425 if (file->isInternal() || !isStaticRelSecType(type))
426 return nullptr;
427 ArrayRef<InputSectionBase *> sections = file->getSections();
428 return sections[info];
429}
430
431template <class ELFT, class RelTy>
432void InputSection::copyRelocations(Ctx &ctx, uint8_t *buf) {
433 if (ctx.arg.relax && !ctx.arg.relocatable &&
434 (ctx.arg.emachine == EM_RISCV || ctx.arg.emachine == EM_LOONGARCH)) {
435 // On LoongArch and RISC-V, relaxation might change relocations: copy
436 // from internal ones that are updated by relaxation.
437 InputSectionBase *sec = getRelocatedSection();
438 copyRelocations<ELFT, RelTy>(
439 ctx, buf,
440 llvm::make_range(x: sec->relocations.begin(), y: sec->relocations.end()));
441 } else {
442 // Convert the raw relocations in the input section into Relocation objects
443 // suitable to be used by copyRelocations below.
444 struct MapRel {
445 Ctx &ctx;
446 const ObjFile<ELFT> &file;
447 Relocation operator()(const RelTy &rel) const {
448 // RelExpr is not used so set to a dummy value.
449 return Relocation{R_NONE, rel.getType(ctx.arg.isMips64EL), rel.r_offset,
450 getAddend<ELFT>(rel), &file.getRelocTargetSym(rel)};
451 }
452 };
453
454 using RawRels = ArrayRef<RelTy>;
455 using MapRelIter =
456 llvm::mapped_iterator<typename RawRels::iterator, MapRel>;
457 auto mapRel = MapRel{ctx, *getFile<ELFT>()};
458 RawRels rawRels = getDataAs<RelTy>();
459 auto rels = llvm::make_range(MapRelIter(rawRels.begin(), mapRel),
460 MapRelIter(rawRels.end(), mapRel));
461 copyRelocations<ELFT, RelTy>(ctx, buf, rels);
462 }
463}
464
465// This is used for -r and --emit-relocs. We can't use memcpy to copy
466// relocations because we need to update symbol table offset and section index
467// for each relocation. So we copy relocations one by one.
468template <class ELFT, class RelTy, class RelIt>
469void InputSection::copyRelocations(Ctx &ctx, uint8_t *buf,
470 llvm::iterator_range<RelIt> rels) {
471 const TargetInfo &target = *ctx.target;
472 InputSectionBase *sec = getRelocatedSection();
473 (void)sec->contentMaybeDecompress(); // uncompress if needed
474
475 for (const Relocation &rel : rels) {
476 RelType type = rel.type;
477 const ObjFile<ELFT> *file = getFile<ELFT>();
478 Symbol &sym = *rel.sym;
479
480 auto *p = reinterpret_cast<typename ELFT::Rela *>(buf);
481 buf += sizeof(RelTy);
482
483 if (RelTy::HasAddend)
484 p->r_addend = rel.addend;
485
486 // Output section VA is zero for -r, so r_offset is an offset within the
487 // section, but for --emit-relocs it is a virtual address.
488 p->r_offset = sec->getVA(offset: rel.offset);
489 p->setSymbolAndType(ctx.in.symTab->getSymbolIndex(sym), type,
490 ctx.arg.isMips64EL);
491
492 if (sym.type == STT_SECTION) {
493 // We combine multiple section symbols into only one per
494 // section. This means we have to update the addend. That is
495 // trivial for Elf_Rela, but for Elf_Rel we have to write to the
496 // section data. We do that by adding to the Relocation vector.
497
498 // .eh_frame is horribly special and can reference discarded sections. To
499 // avoid having to parse and recreate .eh_frame, we just replace any
500 // relocation in it pointing to discarded sections with R_*_NONE, which
501 // hopefully creates a frame that is ignored at runtime. Also, don't warn
502 // on .gcc_except_table and debug sections.
503 //
504 // See the comment in maybeReportUndefined for PPC32 .got2 and PPC64 .toc
505 auto *d = dyn_cast<Defined>(Val: &sym);
506 if (!d) {
507 if (!isDebugSection(sec: *sec) && sec->name != ".eh_frame" &&
508 sec->name != ".gcc_except_table" && sec->name != ".got2" &&
509 sec->name != ".toc") {
510 uint32_t secIdx = cast<Undefined>(Val&: sym).discardedSecIdx;
511 Elf_Shdr_Impl<ELFT> sec = file->template getELFShdrs<ELFT>()[secIdx];
512 Warn(ctx) << "relocation refers to a discarded section: "
513 << CHECK2(file->getObj().getSectionName(sec), file)
514 << "\n>>> referenced by " << getObjMsg(offset: p->r_offset);
515 }
516 p->setSymbolAndType(0, 0, false);
517 continue;
518 }
519 SectionBase *section = d->section;
520 assert(section->isLive());
521
522 int64_t addend = rel.addend;
523 const uint8_t *bufLoc = sec->content().begin() + rel.offset;
524 if (!RelTy::HasAddend)
525 addend = target.getImplicitAddend(buf: bufLoc, type);
526
527 if (ctx.arg.emachine == EM_MIPS &&
528 target.getRelExpr(type, s: sym, loc: bufLoc) == RE_MIPS_GOTREL) {
529 // Some MIPS relocations depend on "gp" value. By default,
530 // this value has 0x7ff0 offset from a .got section. But
531 // relocatable files produced by a compiler or a linker
532 // might redefine this default value and we must use it
533 // for a calculation of the relocation result. When we
534 // generate EXE or DSO it's trivial. Generating a relocatable
535 // output is more difficult case because the linker does
536 // not calculate relocations in this mode and loses
537 // individual "gp" values used by each input object file.
538 // As a workaround we add the "gp" value to the relocation
539 // addend and save it back to the file.
540 addend += sec->getFile<ELFT>()->mipsGp0;
541 }
542
543 if (RelTy::HasAddend)
544 p->r_addend =
545 sym.getVA(ctx, addend) - section->getOutputSection()->addr;
546 // For SHF_ALLOC sections relocated by REL, append a relocation to
547 // sec->relocations so that relocateAlloc transitively called by
548 // writeSections will update the implicit addend. Non-SHF_ALLOC sections
549 // utilize relocateNonAlloc to process raw relocations and do not need
550 // this sec->relocations change.
551 else if (ctx.arg.relocatable && (sec->flags & SHF_ALLOC) &&
552 type != target.noneRel)
553 sec->addReloc(r: {.expr: R_ABS, .type: type, .offset: rel.offset, .addend: addend, .sym: &sym});
554 } else if (ctx.arg.emachine == EM_PPC && type == R_PPC_PLTREL24 &&
555 p->r_addend >= 0x8000 && sec->file->ppc32Got2) {
556 // Similar to R_MIPS_GPREL{16,32}. If the addend of R_PPC_PLTREL24
557 // indicates that r30 is relative to the input section .got2
558 // (r_addend>=0x8000), after linking, r30 should be relative to the output
559 // section .got2 . To compensate for the shift, adjust r_addend by
560 // ppc32Got->outSecOff.
561 p->r_addend += sec->file->ppc32Got2->outSecOff;
562 }
563 }
564}
565
566// The ARM and AArch64 ABI handle pc-relative relocations to undefined weak
567// references specially. The general rule is that the value of the symbol in
568// this context is the address of the place P. A further special case is that
569// branch relocations to an undefined weak reference resolve to the next
570// instruction.
571static uint32_t getARMUndefinedRelativeWeakVA(RelType type, uint32_t a,
572 uint32_t p) {
573 switch (type) {
574 // Unresolved branch relocations to weak references resolve to next
575 // instruction, this will be either 2 or 4 bytes on from P.
576 case R_ARM_THM_JUMP8:
577 case R_ARM_THM_JUMP11:
578 return p + 2 + a;
579 case R_ARM_CALL:
580 case R_ARM_JUMP24:
581 case R_ARM_PC24:
582 case R_ARM_PLT32:
583 case R_ARM_PREL31:
584 case R_ARM_THM_JUMP19:
585 case R_ARM_THM_JUMP24:
586 return p + 4 + a;
587 case R_ARM_THM_CALL:
588 // We don't want an interworking BLX to ARM
589 return p + 5 + a;
590 // Unresolved non branch pc-relative relocations
591 // R_ARM_TARGET2 which can be resolved relatively is not present as it never
592 // targets a weak-reference.
593 case R_ARM_MOVW_PREL_NC:
594 case R_ARM_MOVT_PREL:
595 case R_ARM_REL32:
596 case R_ARM_THM_ALU_PREL_11_0:
597 case R_ARM_THM_MOVW_PREL_NC:
598 case R_ARM_THM_MOVT_PREL:
599 case R_ARM_THM_PC12:
600 return p + a;
601 // p + a is unrepresentable as negative immediates can't be encoded.
602 case R_ARM_THM_PC8:
603 return p;
604 }
605 llvm_unreachable("ARM pc-relative relocation expected\n");
606}
607
608// The comment above getARMUndefinedRelativeWeakVA applies to this function.
609static uint64_t getAArch64UndefinedRelativeWeakVA(uint64_t type, uint64_t p) {
610 switch (type) {
611 // Unresolved branch relocations to weak references resolve to next
612 // instruction, this is 4 bytes on from P.
613 case R_AARCH64_CALL26:
614 case R_AARCH64_CONDBR19:
615 case R_AARCH64_JUMP26:
616 case R_AARCH64_TSTBR14:
617 return p + 4;
618 // Unresolved non branch pc-relative relocations
619 case R_AARCH64_PREL16:
620 case R_AARCH64_PREL32:
621 case R_AARCH64_PREL64:
622 case R_AARCH64_ADR_PREL_LO21:
623 case R_AARCH64_LD_PREL_LO19:
624 case R_AARCH64_PLT32:
625 return p;
626 }
627 llvm_unreachable("AArch64 pc-relative relocation expected\n");
628}
629
630static uint64_t getRISCVUndefinedRelativeWeakVA(uint64_t type, uint64_t p) {
631 switch (type) {
632 case R_RISCV_BRANCH:
633 case R_RISCV_JAL:
634 case R_RISCV_CALL:
635 case R_RISCV_CALL_PLT:
636 case R_RISCV_RVC_BRANCH:
637 case R_RISCV_RVC_JUMP:
638 case R_RISCV_PLT32:
639 return p;
640 default:
641 return 0;
642 }
643}
644
645// ARM SBREL relocations are of the form S + A - B where B is the static base
646// The ARM ABI defines base to be "addressing origin of the output segment
647// defining the symbol S". We defined the "addressing origin"/static base to be
648// the base of the PT_LOAD segment containing the Sym.
649// The procedure call standard only defines a Read Write Position Independent
650// RWPI variant so in practice we should expect the static base to be the base
651// of the RW segment.
652static uint64_t getARMStaticBase(const Symbol &sym) {
653 OutputSection *os = sym.getOutputSection();
654 if (!os || !os->ptLoad || !os->ptLoad->firstSec) {
655 Err(ctx&: os->ctx) << "SBREL relocation to " << sym.getName()
656 << " without static base";
657 return 0;
658 }
659 return os->ptLoad->firstSec->addr;
660}
661
662// For RE_RISCV_PC_INDIRECT (R_RISCV_PCREL_LO12_{I,S}), the symbol actually
663// points the corresponding R_RISCV_PCREL_HI20 relocation, and the target VA
664// is calculated using PCREL_HI20's symbol.
665//
666// This function returns the R_RISCV_PCREL_HI20 relocation from the
667// R_RISCV_PCREL_LO12 relocation.
668static Relocation *getRISCVPCRelHi20(Ctx &ctx, const InputSectionBase *loSec,
669 const Relocation &loReloc) {
670 uint64_t addend = loReloc.addend;
671 Symbol *sym = loReloc.sym;
672
673 const Defined *d = cast<Defined>(Val: sym);
674 if (!d->section) {
675 Err(ctx) << loSec->getLocation(offset: loReloc.offset)
676 << ": R_RISCV_PCREL_LO12 relocation points to an absolute symbol: "
677 << sym->getName();
678 return nullptr;
679 }
680 InputSection *hiSec = cast<InputSection>(Val: d->section);
681
682 if (hiSec != loSec)
683 Err(ctx) << loSec->getLocation(offset: loReloc.offset)
684 << ": R_RISCV_PCREL_LO12 relocation points to a symbol '"
685 << sym->getName() << "' in a different section '" << hiSec->name
686 << "'";
687
688 if (addend != 0)
689 Warn(ctx) << loSec->getLocation(offset: loReloc.offset)
690 << ": non-zero addend in R_RISCV_PCREL_LO12 relocation to "
691 << hiSec->getObjMsg(offset: d->value) << " is ignored";
692
693 // Relocations are sorted by offset, so we can use std::equal_range to do
694 // binary search.
695 Relocation hiReloc;
696 hiReloc.offset = d->value;
697 auto range =
698 std::equal_range(first: hiSec->relocs().begin(), last: hiSec->relocs().end(), val: hiReloc,
699 comp: [](const Relocation &lhs, const Relocation &rhs) {
700 return lhs.offset < rhs.offset;
701 });
702
703 for (auto it = range.first; it != range.second; ++it)
704 if (it->type == R_RISCV_PCREL_HI20 || it->type == R_RISCV_GOT_HI20 ||
705 it->type == R_RISCV_TLS_GD_HI20 || it->type == R_RISCV_TLS_GOT_HI20)
706 return &*it;
707
708 Err(ctx) << loSec->getLocation(offset: loReloc.offset)
709 << ": R_RISCV_PCREL_LO12 relocation points to "
710 << hiSec->getObjMsg(offset: d->value)
711 << " without an associated R_RISCV_PCREL_HI20 relocation";
712 return nullptr;
713}
714
715// A TLS symbol's virtual address is relative to the TLS segment. Add a
716// target-specific adjustment to produce a thread-pointer-relative offset.
717static int64_t getTlsTpOffset(Ctx &ctx, const Symbol &s) {
718 // On targets that support TLSDESC, _TLS_MODULE_BASE_@tpoff = 0.
719 if (&s == ctx.sym.tlsModuleBase)
720 return 0;
721
722 // There are 2 TLS layouts. Among targets we support, x86 uses TLS Variant 2
723 // while most others use Variant 1. At run time TP will be aligned to p_align.
724
725 // Variant 1. TP will be followed by an optional gap (which is the size of 2
726 // pointers on ARM/AArch64, 0 on other targets), followed by alignment
727 // padding, then the static TLS blocks. The alignment padding is added so that
728 // (TP + gap + padding) is congruent to p_vaddr modulo p_align.
729 //
730 // Variant 2. Static TLS blocks, followed by alignment padding are placed
731 // before TP. The alignment padding is added so that (TP - padding -
732 // p_memsz) is congruent to p_vaddr modulo p_align.
733 PhdrEntry *tls = ctx.tlsPhdr;
734 if (!tls) // Reported an error in getSymVA
735 return 0;
736 switch (ctx.arg.emachine) {
737 // Variant 1.
738 case EM_ARM:
739 case EM_AARCH64:
740 return s.getVA(ctx, addend: 0) + ctx.arg.wordsize * 2 +
741 ((tls->p_vaddr - ctx.arg.wordsize * 2) & (tls->p_align - 1));
742 case EM_MIPS:
743 case EM_PPC:
744 case EM_PPC64:
745 // Adjusted Variant 1. TP is placed with a displacement of 0x7000, which is
746 // to allow a signed 16-bit offset to reach 0x1000 of TCB/thread-library
747 // data and 0xf000 of the program's TLS segment.
748 return s.getVA(ctx, addend: 0) + (tls->p_vaddr & (tls->p_align - 1)) - 0x7000;
749 case EM_LOONGARCH:
750 case EM_RISCV:
751 // See the comment in handleTlsRelocation. For TLSDESC=>IE,
752 // R_RISCV_TLSDESC_{LOAD_LO12,ADD_LO12_I,CALL} also reach here. While
753 // `tls` may be null, the return value is ignored.
754 if (s.type != STT_TLS)
755 return 0;
756 return s.getVA(ctx, addend: 0) + (tls->p_vaddr & (tls->p_align - 1));
757
758 // Variant 2.
759 case EM_HEXAGON:
760 case EM_S390:
761 case EM_SPARCV9:
762 case EM_386:
763 case EM_X86_64:
764 return s.getVA(ctx, addend: 0) - tls->p_memsz -
765 ((-tls->p_vaddr - tls->p_memsz) & (tls->p_align - 1));
766 default:
767 llvm_unreachable("unhandled ctx.arg.emachine");
768 }
769}
770
771uint64_t InputSectionBase::getRelocTargetVA(Ctx &ctx, const Relocation &r,
772 uint64_t p) const {
773 int64_t a = r.addend;
774 switch (r.expr) {
775 case R_ABS:
776 case R_DTPREL:
777 case R_RELAX_TLS_LD_TO_LE_ABS:
778 case R_RELAX_GOT_PC_NOPIC:
779 case RE_AARCH64_AUTH:
780 case RE_RISCV_ADD:
781 case RE_RISCV_LEB128:
782 return r.sym->getVA(ctx, addend: a);
783 case R_ADDEND:
784 return a;
785 case R_RELAX_HINT:
786 return 0;
787 case RE_ARM_SBREL:
788 return r.sym->getVA(ctx, addend: a) - getARMStaticBase(sym: *r.sym);
789 case R_GOT:
790 case RE_AARCH64_AUTH_GOT:
791 case R_RELAX_TLS_GD_TO_IE_ABS:
792 return r.sym->getGotVA(ctx) + a;
793 case RE_LOONGARCH_GOT:
794 // The LoongArch TLS GD relocs reuse the R_LARCH_GOT_PC_LO12 reloc r.type
795 // for their page offsets. The arithmetics are different in the TLS case
796 // so we have to duplicate some logic here.
797 if (r.sym->hasFlag(bit: NEEDS_TLSGD) && r.type != R_LARCH_TLS_IE_PC_LO12)
798 // Like RE_LOONGARCH_TLSGD_PAGE_PC but taking the absolute value.
799 return ctx.in.got->getGlobalDynAddr(b: *r.sym) + a;
800 return r.sym->getGotVA(ctx) + a;
801 case R_GOTONLY_PC:
802 return ctx.in.got->getVA() + a - p;
803 case R_GOTPLTONLY_PC:
804 return ctx.in.gotPlt->getVA() + a - p;
805 case R_GOTREL:
806 case RE_PPC64_RELAX_TOC:
807 return r.sym->getVA(ctx, addend: a) - ctx.in.got->getVA();
808 case R_GOTPLTREL:
809 return r.sym->getVA(ctx, addend: a) - ctx.in.gotPlt->getVA();
810 case R_GOTPLT:
811 case R_RELAX_TLS_GD_TO_IE_GOTPLT:
812 return r.sym->getGotVA(ctx) + a - ctx.in.gotPlt->getVA();
813 case R_TLSLD_GOT_OFF:
814 case R_GOT_OFF:
815 case R_RELAX_TLS_GD_TO_IE_GOT_OFF:
816 return r.sym->getGotOffset(ctx) + a;
817 case RE_AARCH64_GOT_PAGE_PC:
818 case RE_AARCH64_AUTH_GOT_PAGE_PC:
819 case RE_AARCH64_RELAX_TLS_GD_TO_IE_PAGE_PC:
820 return getAArch64Page(expr: r.sym->getGotVA(ctx) + a) - getAArch64Page(expr: p);
821 case RE_AARCH64_GOT_PAGE:
822 return r.sym->getGotVA(ctx) + a - getAArch64Page(expr: ctx.in.got->getVA());
823 case R_GOT_PC:
824 case RE_AARCH64_AUTH_GOT_PC:
825 case R_RELAX_TLS_GD_TO_IE:
826 return r.sym->getGotVA(ctx) + a - p;
827 case R_GOTPLT_GOTREL:
828 return r.sym->getGotPltVA(ctx) + a - ctx.in.got->getVA();
829 case R_GOTPLT_PC:
830 return r.sym->getGotPltVA(ctx) + a - p;
831 case RE_LOONGARCH_GOT_PAGE_PC:
832 if (r.sym->hasFlag(bit: NEEDS_TLSGD))
833 return getLoongArchPageDelta(dest: ctx.in.got->getGlobalDynAddr(b: *r.sym) + a, pc: p,
834 type: r.type);
835 return getLoongArchPageDelta(dest: r.sym->getGotVA(ctx) + a, pc: p, type: r.type);
836 case RE_MIPS_GOTREL:
837 return r.sym->getVA(ctx, addend: a) - ctx.in.mipsGot->getGp(f: file);
838 case RE_MIPS_GOT_GP:
839 return ctx.in.mipsGot->getGp(f: file) + a;
840 case RE_MIPS_GOT_GP_PC: {
841 // R_MIPS_LO16 expression has RE_MIPS_GOT_GP_PC r.type iif the target
842 // is _gp_disp symbol. In that case we should use the following
843 // formula for calculation "AHL + GP - P + 4". For details see p. 4-19 at
844 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
845 // microMIPS variants of these relocations use slightly different
846 // expressions: AHL + GP - P + 3 for %lo() and AHL + GP - P - 1 for %hi()
847 // to correctly handle less-significant bit of the microMIPS symbol.
848 uint64_t v = ctx.in.mipsGot->getGp(f: file) + a - p;
849 if (r.type == R_MIPS_LO16 || r.type == R_MICROMIPS_LO16)
850 v += 4;
851 if (r.type == R_MICROMIPS_LO16 || r.type == R_MICROMIPS_HI16)
852 v -= 1;
853 return v;
854 }
855 case RE_MIPS_GOT_LOCAL_PAGE:
856 // If relocation against MIPS local symbol requires GOT entry, this entry
857 // should be initialized by 'page address'. This address is high 16-bits
858 // of sum the symbol's value and the addend.
859 return ctx.in.mipsGot->getVA() +
860 ctx.in.mipsGot->getPageEntryOffset(f: file, s: *r.sym, addend: a) -
861 ctx.in.mipsGot->getGp(f: file);
862 case RE_MIPS_GOT_OFF:
863 case RE_MIPS_GOT_OFF32:
864 // In case of MIPS if a GOT relocation has non-zero addend this addend
865 // should be applied to the GOT entry content not to the GOT entry offset.
866 // That is why we use separate expression r.type.
867 return ctx.in.mipsGot->getVA() +
868 ctx.in.mipsGot->getSymEntryOffset(f: file, s: *r.sym, addend: a) -
869 ctx.in.mipsGot->getGp(f: file);
870 case RE_MIPS_TLSGD:
871 return ctx.in.mipsGot->getVA() +
872 ctx.in.mipsGot->getGlobalDynOffset(f: file, s: *r.sym) -
873 ctx.in.mipsGot->getGp(f: file);
874 case RE_MIPS_TLSLD:
875 return ctx.in.mipsGot->getVA() + ctx.in.mipsGot->getTlsIndexOffset(f: file) -
876 ctx.in.mipsGot->getGp(f: file);
877 case RE_AARCH64_PAGE_PC: {
878 uint64_t val = r.sym->isUndefWeak() ? p + a : r.sym->getVA(ctx, addend: a);
879 return getAArch64Page(expr: val) - getAArch64Page(expr: p);
880 }
881 case RE_RISCV_PC_INDIRECT: {
882 if (const Relocation *hiRel = getRISCVPCRelHi20(ctx, loSec: this, loReloc: r))
883 return getRelocTargetVA(ctx, r: *hiRel, p: r.sym->getVA(ctx));
884 return 0;
885 }
886 case RE_LOONGARCH_PAGE_PC:
887 return getLoongArchPageDelta(dest: r.sym->getVA(ctx, addend: a), pc: p, type: r.type);
888 case R_PC:
889 case RE_ARM_PCA: {
890 uint64_t dest;
891 if (r.expr == RE_ARM_PCA)
892 // Some PC relative ARM (Thumb) relocations align down the place.
893 p = p & 0xfffffffc;
894 if (r.sym->isUndefined()) {
895 // On ARM and AArch64 a branch to an undefined weak resolves to the next
896 // instruction, otherwise the place. On RISC-V, resolve an undefined weak
897 // to the same instruction to cause an infinite loop (making the user
898 // aware of the issue) while ensuring no overflow.
899 // Note: if the symbol is hidden, its binding has been converted to local,
900 // so we just check isUndefined() here.
901 if (ctx.arg.emachine == EM_ARM)
902 dest = getARMUndefinedRelativeWeakVA(type: r.type, a, p);
903 else if (ctx.arg.emachine == EM_AARCH64)
904 dest = getAArch64UndefinedRelativeWeakVA(type: r.type, p) + a;
905 else if (ctx.arg.emachine == EM_PPC)
906 dest = p;
907 else if (ctx.arg.emachine == EM_RISCV)
908 dest = getRISCVUndefinedRelativeWeakVA(type: r.type, p) + a;
909 else
910 dest = r.sym->getVA(ctx, addend: a);
911 } else {
912 dest = r.sym->getVA(ctx, addend: a);
913 }
914 return dest - p;
915 }
916 case R_PLT:
917 return r.sym->getPltVA(ctx) + a;
918 case R_PLT_PC:
919 case RE_PPC64_CALL_PLT:
920 return r.sym->getPltVA(ctx) + a - p;
921 case RE_LOONGARCH_PLT_PAGE_PC:
922 return getLoongArchPageDelta(dest: r.sym->getPltVA(ctx) + a, pc: p, type: r.type);
923 case R_PLT_GOTPLT:
924 return r.sym->getPltVA(ctx) + a - ctx.in.gotPlt->getVA();
925 case R_PLT_GOTREL:
926 return r.sym->getPltVA(ctx) + a - ctx.in.got->getVA();
927 case RE_PPC32_PLTREL:
928 // R_PPC_PLTREL24 uses the addend (usually 0 or 0x8000) to indicate r30
929 // stores _GLOBAL_OFFSET_TABLE_ or .got2+0x8000. The addend is ignored for
930 // target VA computation.
931 return r.sym->getPltVA(ctx) - p;
932 case RE_PPC64_CALL: {
933 uint64_t symVA = r.sym->getVA(ctx, addend: a);
934 // If we have an undefined weak symbol, we might get here with a symbol
935 // address of zero. That could overflow, but the code must be unreachable,
936 // so don't bother doing anything at all.
937 if (!symVA)
938 return 0;
939
940 // PPC64 V2 ABI describes two entry points to a function. The global entry
941 // point is used for calls where the caller and callee (may) have different
942 // TOC base pointers and r2 needs to be modified to hold the TOC base for
943 // the callee. For local calls the caller and callee share the same
944 // TOC base and so the TOC pointer initialization code should be skipped by
945 // branching to the local entry point.
946 return symVA - p +
947 getPPC64GlobalEntryToLocalEntryOffset(ctx, stOther: r.sym->stOther);
948 }
949 case RE_PPC64_TOCBASE:
950 return getPPC64TocBase(ctx) + a;
951 case R_RELAX_GOT_PC:
952 case RE_PPC64_RELAX_GOT_PC:
953 return r.sym->getVA(ctx, addend: a) - p;
954 case R_RELAX_TLS_GD_TO_LE:
955 case R_RELAX_TLS_IE_TO_LE:
956 case R_RELAX_TLS_LD_TO_LE:
957 case R_TPREL:
958 // It is not very clear what to return if the symbol is undefined. With
959 // --noinhibit-exec, even a non-weak undefined reference may reach here.
960 // Just return A, which matches R_ABS, and the behavior of some dynamic
961 // loaders.
962 if (r.sym->isUndefined())
963 return a;
964 return getTlsTpOffset(ctx, s: *r.sym) + a;
965 case R_RELAX_TLS_GD_TO_LE_NEG:
966 case R_TPREL_NEG:
967 if (r.sym->isUndefined())
968 return a;
969 return -getTlsTpOffset(ctx, s: *r.sym) + a;
970 case R_SIZE:
971 return r.sym->getSize() + a;
972 case R_TLSDESC:
973 case RE_AARCH64_AUTH_TLSDESC:
974 return ctx.in.got->getTlsDescAddr(sym: *r.sym) + a;
975 case R_TLSDESC_PC:
976 return ctx.in.got->getTlsDescAddr(sym: *r.sym) + a - p;
977 case R_TLSDESC_GOTPLT:
978 return ctx.in.got->getTlsDescAddr(sym: *r.sym) + a - ctx.in.gotPlt->getVA();
979 case RE_AARCH64_TLSDESC_PAGE:
980 case RE_AARCH64_AUTH_TLSDESC_PAGE:
981 return getAArch64Page(expr: ctx.in.got->getTlsDescAddr(sym: *r.sym) + a) -
982 getAArch64Page(expr: p);
983 case RE_LOONGARCH_TLSDESC_PAGE_PC:
984 return getLoongArchPageDelta(dest: ctx.in.got->getTlsDescAddr(sym: *r.sym) + a, pc: p,
985 type: r.type);
986 case R_TLSGD_GOT:
987 return ctx.in.got->getGlobalDynOffset(b: *r.sym) + a;
988 case R_TLSGD_GOTPLT:
989 return ctx.in.got->getGlobalDynAddr(b: *r.sym) + a - ctx.in.gotPlt->getVA();
990 case R_TLSGD_PC:
991 return ctx.in.got->getGlobalDynAddr(b: *r.sym) + a - p;
992 case RE_LOONGARCH_TLSGD_PAGE_PC:
993 return getLoongArchPageDelta(dest: ctx.in.got->getGlobalDynAddr(b: *r.sym) + a, pc: p,
994 type: r.type);
995 case R_TLSLD_GOTPLT:
996 return ctx.in.got->getVA() + ctx.in.got->getTlsIndexOff() + a -
997 ctx.in.gotPlt->getVA();
998 case R_TLSLD_GOT:
999 return ctx.in.got->getTlsIndexOff() + a;
1000 case R_TLSLD_PC:
1001 return ctx.in.got->getTlsIndexVA() + a - p;
1002 default:
1003 llvm_unreachable("invalid expression");
1004 }
1005}
1006
1007// This function applies relocations to sections without SHF_ALLOC bit.
1008// Such sections are never mapped to memory at runtime. Debug sections are
1009// an example. Relocations in non-alloc sections are much easier to
1010// handle than in allocated sections because it will never need complex
1011// treatment such as GOT or PLT (because at runtime no one refers them).
1012// So, we handle relocations for non-alloc sections directly in this
1013// function as a performance optimization.
1014template <class ELFT, class RelTy>
1015void InputSection::relocateNonAlloc(Ctx &ctx, uint8_t *buf,
1016 Relocs<RelTy> rels) {
1017 const unsigned bits = sizeof(typename ELFT::uint) * 8;
1018 const TargetInfo &target = *ctx.target;
1019 const auto emachine = ctx.arg.emachine;
1020 const bool isDebug = isDebugSection(sec: *this);
1021 const bool isDebugLine = isDebug && name == ".debug_line";
1022 std::optional<uint64_t> tombstone;
1023 if (isDebug) {
1024 if (name == ".debug_loc" || name == ".debug_ranges")
1025 tombstone = 1;
1026 else if (name == ".debug_names")
1027 tombstone = UINT64_MAX; // tombstone value
1028 else
1029 tombstone = 0;
1030 }
1031 for (const auto &patAndValue : llvm::reverse(C&: ctx.arg.deadRelocInNonAlloc))
1032 if (patAndValue.first.match(S: this->name)) {
1033 tombstone = patAndValue.second;
1034 break;
1035 }
1036
1037 const InputFile *f = this->file;
1038 for (auto it = rels.begin(), end = rels.end(); it != end; ++it) {
1039 const RelTy &rel = *it;
1040 const RelType type = rel.getType(ctx.arg.isMips64EL);
1041 const uint64_t offset = rel.r_offset;
1042 uint8_t *bufLoc = buf + offset;
1043 int64_t addend = getAddend<ELFT>(rel);
1044 if (!RelTy::HasAddend)
1045 addend += target.getImplicitAddend(buf: bufLoc, type);
1046
1047 Symbol &sym = f->getRelocTargetSym(rel);
1048 RelExpr expr = target.getRelExpr(type, s: sym, loc: bufLoc);
1049 if (expr == R_NONE)
1050 continue;
1051 auto *ds = dyn_cast<Defined>(Val: &sym);
1052
1053 if (emachine == EM_RISCV && type == R_RISCV_SET_ULEB128) {
1054 if (++it != end &&
1055 it->getType(/*isMips64EL=*/false) == R_RISCV_SUB_ULEB128 &&
1056 it->r_offset == offset) {
1057 uint64_t val;
1058 if (!ds && tombstone) {
1059 val = *tombstone;
1060 } else {
1061 val = sym.getVA(ctx, addend) -
1062 (f->getRelocTargetSym(*it).getVA(ctx) + getAddend<ELFT>(*it));
1063 }
1064 if (overwriteULEB128(bufLoc, val) >= 0x80)
1065 Err(ctx) << getLocation(offset) << ": ULEB128 value " << val
1066 << " exceeds available space; references '" << &sym << "'";
1067 continue;
1068 }
1069 Err(ctx) << getLocation(offset)
1070 << ": R_RISCV_SET_ULEB128 not paired with R_RISCV_SUB_SET128";
1071 return;
1072 }
1073
1074 if (tombstone && (expr == R_ABS || expr == R_DTPREL)) {
1075 // Resolve relocations in .debug_* referencing (discarded symbols or ICF
1076 // folded section symbols) to a tombstone value. Resolving to addend is
1077 // unsatisfactory because the result address range may collide with a
1078 // valid range of low address, or leave multiple CUs claiming ownership of
1079 // the same range of code, which may confuse consumers.
1080 //
1081 // To address the problems, we use -1 as a tombstone value for most
1082 // .debug_* sections. We have to ignore the addend because we don't want
1083 // to resolve an address attribute (which may have a non-zero addend) to
1084 // -1+addend (wrap around to a low address).
1085 //
1086 // R_DTPREL type relocations represent an offset into the dynamic thread
1087 // vector. The computed value is st_value plus a non-negative offset.
1088 // Negative values are invalid, so -1 can be used as the tombstone value.
1089 //
1090 // If the referenced symbol is relative to a discarded section (due to
1091 // --gc-sections, COMDAT, etc), it has been converted to a Undefined.
1092 // `ds->folded` catches the ICF folded case. However, resolving a
1093 // relocation in .debug_line to -1 would stop debugger users from setting
1094 // breakpoints on the folded-in function, so exclude .debug_line.
1095 //
1096 // For pre-DWARF-v5 .debug_loc and .debug_ranges, -1 is a reserved value
1097 // (base address selection entry), use 1 (which is used by GNU ld for
1098 // .debug_ranges).
1099 //
1100 // TODO To reduce disruption, we use 0 instead of -1 as the tombstone
1101 // value. Enable -1 in a future release.
1102 if (!ds || (ds->folded && !isDebugLine)) {
1103 // If -z dead-reloc-in-nonalloc= is specified, respect it.
1104 uint64_t value = SignExtend64<bits>(*tombstone);
1105 // For a 32-bit local TU reference in .debug_names, X86_64::relocate
1106 // requires that the unsigned value for R_X86_64_32 is truncated to
1107 // 32-bit. Other 64-bit targets's don't discern signed/unsigned 32-bit
1108 // absolute relocations and do not need this change.
1109 if (emachine == EM_X86_64 && type == R_X86_64_32)
1110 value = static_cast<uint32_t>(value);
1111 target.relocateNoSym(loc: bufLoc, type, val: value);
1112 continue;
1113 }
1114 }
1115
1116 // For a relocatable link, content relocated by relocation types with an
1117 // explicit addend, such as RELA, remain unchanged and we can stop here.
1118 // While content relocated by relocation types with an implicit addend, such
1119 // as REL, needs the implicit addend updated.
1120 if (ctx.arg.relocatable && (RelTy::HasAddend || sym.type != STT_SECTION))
1121 continue;
1122
1123 // R_ABS/R_DTPREL and some other relocations can be used from non-SHF_ALLOC
1124 // sections.
1125 if (LLVM_LIKELY(expr == R_ABS) || expr == R_DTPREL || expr == R_GOTPLTREL ||
1126 expr == RE_RISCV_ADD || expr == RE_ARM_SBREL) {
1127 target.relocateNoSym(loc: bufLoc, type,
1128 val: SignExtend64<bits>(sym.getVA(ctx, addend)));
1129 continue;
1130 }
1131
1132 if (expr == R_SIZE) {
1133 target.relocateNoSym(loc: bufLoc, type,
1134 val: SignExtend64<bits>(sym.getSize() + addend));
1135 continue;
1136 }
1137
1138 // If the control reaches here, we found a PC-relative relocation in a
1139 // non-ALLOC section. Since non-ALLOC section is not loaded into memory
1140 // at runtime, the notion of PC-relative doesn't make sense here. So,
1141 // this is a usage error. However, GNU linkers historically accept such
1142 // relocations without any errors and relocate them as if they were at
1143 // address 0. For bug-compatibility, we accept them with warnings. We
1144 // know Steel Bank Common Lisp as of 2018 have this bug.
1145 //
1146 // GCC 8.0 or earlier have a bug that they emit R_386_GOTPC relocations
1147 // against _GLOBAL_OFFSET_TABLE_ for .debug_info. The bug has been fixed in
1148 // 2017 (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82630), but we need to
1149 // keep this bug-compatible code for a while.
1150 bool isErr = expr != R_PC && !(emachine == EM_386 && type == R_386_GOTPC);
1151 {
1152 ELFSyncStream diag(ctx, isErr && !ctx.arg.noinhibitExec
1153 ? DiagLevel::Err
1154 : DiagLevel::Warn);
1155 diag << getLocation(offset) << ": has non-ABS relocation " << type
1156 << " against symbol '" << &sym << "'";
1157 }
1158 if (!isErr)
1159 target.relocateNoSym(
1160 loc: bufLoc, type,
1161 val: SignExtend64<bits>(sym.getVA(ctx, addend: addend - offset - outSecOff)));
1162 }
1163}
1164
1165template <class ELFT>
1166void InputSectionBase::relocate(Ctx &ctx, uint8_t *buf, uint8_t *bufEnd) {
1167 if ((flags & SHF_EXECINSTR) && LLVM_UNLIKELY(getFile<ELFT>()->splitStack))
1168 adjustSplitStackFunctionPrologues<ELFT>(ctx, buf, bufEnd);
1169
1170 if (flags & SHF_ALLOC) {
1171 ctx.target->relocateAlloc(sec&: *this, buf);
1172 return;
1173 }
1174
1175 auto *sec = cast<InputSection>(Val: this);
1176 // For a relocatable link, also call relocateNonAlloc() to rewrite applicable
1177 // locations with tombstone values.
1178 invokeOnRelocs(*sec, sec->relocateNonAlloc<ELFT>, ctx, buf);
1179}
1180
1181// For each function-defining prologue, find any calls to __morestack,
1182// and replace them with calls to __morestack_non_split.
1183static void switchMorestackCallsToMorestackNonSplit(
1184 Ctx &ctx, DenseSet<Defined *> &prologues,
1185 SmallVector<Relocation *, 0> &morestackCalls) {
1186
1187 // If the target adjusted a function's prologue, all calls to
1188 // __morestack inside that function should be switched to
1189 // __morestack_non_split.
1190 Symbol *moreStackNonSplit = ctx.symtab->find(name: "__morestack_non_split");
1191 if (!moreStackNonSplit) {
1192 ErrAlways(ctx) << "mixing split-stack objects requires a definition of "
1193 "__morestack_non_split";
1194 return;
1195 }
1196
1197 // Sort both collections to compare addresses efficiently.
1198 llvm::sort(C&: morestackCalls, Comp: [](const Relocation *l, const Relocation *r) {
1199 return l->offset < r->offset;
1200 });
1201 std::vector<Defined *> functions(prologues.begin(), prologues.end());
1202 llvm::sort(C&: functions, Comp: [](const Defined *l, const Defined *r) {
1203 return l->value < r->value;
1204 });
1205
1206 auto it = morestackCalls.begin();
1207 for (Defined *f : functions) {
1208 // Find the first call to __morestack within the function.
1209 while (it != morestackCalls.end() && (*it)->offset < f->value)
1210 ++it;
1211 // Adjust all calls inside the function.
1212 while (it != morestackCalls.end() && (*it)->offset < f->value + f->size) {
1213 (*it)->sym = moreStackNonSplit;
1214 ++it;
1215 }
1216 }
1217}
1218
1219static bool enclosingPrologueAttempted(uint64_t offset,
1220 const DenseSet<Defined *> &prologues) {
1221 for (Defined *f : prologues)
1222 if (f->value <= offset && offset < f->value + f->size)
1223 return true;
1224 return false;
1225}
1226
1227// If a function compiled for split stack calls a function not
1228// compiled for split stack, then the caller needs its prologue
1229// adjusted to ensure that the called function will have enough stack
1230// available. Find those functions, and adjust their prologues.
1231template <class ELFT>
1232void InputSectionBase::adjustSplitStackFunctionPrologues(Ctx &ctx, uint8_t *buf,
1233 uint8_t *end) {
1234 DenseSet<Defined *> prologues;
1235 SmallVector<Relocation *, 0> morestackCalls;
1236
1237 for (Relocation &rel : relocs()) {
1238 // Ignore calls into the split-stack api.
1239 if (rel.sym->getName().starts_with(Prefix: "__morestack")) {
1240 if (rel.sym->getName() == "__morestack")
1241 morestackCalls.push_back(Elt: &rel);
1242 continue;
1243 }
1244
1245 // A relocation to non-function isn't relevant. Sometimes
1246 // __morestack is not marked as a function, so this check comes
1247 // after the name check.
1248 if (rel.sym->type != STT_FUNC)
1249 continue;
1250
1251 // If the callee's-file was compiled with split stack, nothing to do. In
1252 // this context, a "Defined" symbol is one "defined by the binary currently
1253 // being produced". So an "undefined" symbol might be provided by a shared
1254 // library. It is not possible to tell how such symbols were compiled, so be
1255 // conservative.
1256 if (Defined *d = dyn_cast<Defined>(Val: rel.sym))
1257 if (InputSection *isec = cast_or_null<InputSection>(Val: d->section))
1258 if (!isec || !isec->getFile<ELFT>() || isec->getFile<ELFT>()->splitStack)
1259 continue;
1260
1261 if (enclosingPrologueAttempted(offset: rel.offset, prologues))
1262 continue;
1263
1264 if (Defined *f = getEnclosingFunction(offset: rel.offset)) {
1265 prologues.insert(V: f);
1266 if (ctx.target->adjustPrologueForCrossSplitStack(loc: buf + f->value, end,
1267 stOther: f->stOther))
1268 continue;
1269 if (!getFile<ELFT>()->someNoSplitStack)
1270 Err(ctx)
1271 << this << ": " << f->getName() << " (with -fsplit-stack) calls "
1272 << rel.sym->getName()
1273 << " (without -fsplit-stack), but couldn't adjust its prologue";
1274 }
1275 }
1276
1277 if (ctx.target->needsMoreStackNonSplit)
1278 switchMorestackCallsToMorestackNonSplit(ctx, prologues, morestackCalls);
1279}
1280
1281template <class ELFT> void InputSection::writeTo(Ctx &ctx, uint8_t *buf) {
1282 if (LLVM_UNLIKELY(type == SHT_NOBITS))
1283 return;
1284 // If -r or --emit-relocs is given, then an InputSection
1285 // may be a relocation section.
1286 if (LLVM_UNLIKELY(type == SHT_RELA)) {
1287 copyRelocations<ELFT, typename ELFT::Rela>(ctx, buf);
1288 return;
1289 }
1290 if (LLVM_UNLIKELY(type == SHT_REL)) {
1291 copyRelocations<ELFT, typename ELFT::Rel>(ctx, buf);
1292 return;
1293 }
1294
1295 // If -r is given, we may have a SHT_GROUP section.
1296 if (LLVM_UNLIKELY(type == SHT_GROUP)) {
1297 copyShtGroup<ELFT>(buf);
1298 return;
1299 }
1300
1301 // If this is a compressed section, uncompress section contents directly
1302 // to the buffer.
1303 if (compressed) {
1304 auto *hdr = reinterpret_cast<const typename ELFT::Chdr *>(content_);
1305 auto compressed = ArrayRef<uint8_t>(content_, compressedSize)
1306 .slice(N: sizeof(typename ELFT::Chdr));
1307 size_t size = this->size;
1308 if (Error e = hdr->ch_type == ELFCOMPRESS_ZLIB
1309 ? compression::zlib::decompress(Input: compressed, Output: buf, UncompressedSize&: size)
1310 : compression::zstd::decompress(Input: compressed, Output: buf, UncompressedSize&: size))
1311 Err(ctx) << this << ": decompress failed: " << std::move(e);
1312 uint8_t *bufEnd = buf + size;
1313 relocate<ELFT>(ctx, buf, bufEnd);
1314 return;
1315 }
1316
1317 // Copy section contents from source object file to output file
1318 // and then apply relocations.
1319 memcpy(dest: buf, src: content().data(), n: content().size());
1320 relocate<ELFT>(ctx, buf, buf + content().size());
1321}
1322
1323void InputSection::replace(InputSection *other) {
1324 addralign = std::max(a: addralign, b: other->addralign);
1325
1326 // When a section is replaced with another section that was allocated to
1327 // another partition, the replacement section (and its associated sections)
1328 // need to be placed in the main partition so that both partitions will be
1329 // able to access it.
1330 if (partition != other->partition) {
1331 partition = 1;
1332 for (InputSection *isec : dependentSections)
1333 isec->partition = 1;
1334 }
1335
1336 other->repl = repl;
1337 other->markDead();
1338}
1339
1340template <class ELFT>
1341EhInputSection::EhInputSection(ObjFile<ELFT> &f,
1342 const typename ELFT::Shdr &header,
1343 StringRef name)
1344 : InputSectionBase(f, header, name, InputSectionBase::EHFrame) {}
1345
1346SyntheticSection *EhInputSection::getParent() const {
1347 return cast_or_null<SyntheticSection>(Val: parent);
1348}
1349
1350// .eh_frame is a sequence of CIE or FDE records.
1351// This function splits an input section into records and returns them.
1352template <class ELFT> void EhInputSection::split() {
1353 const RelsOrRelas<ELFT> rels = relsOrRelas<ELFT>(/*supportsCrel=*/false);
1354 // getReloc expects the relocations to be sorted by r_offset. See the comment
1355 // in scanRelocs.
1356 if (rels.areRelocsRel()) {
1357 SmallVector<typename ELFT::Rel, 0> storage;
1358 split<ELFT>(sortRels(rels.rels, storage));
1359 } else {
1360 SmallVector<typename ELFT::Rela, 0> storage;
1361 split<ELFT>(sortRels(rels.relas, storage));
1362 }
1363}
1364
1365template <class ELFT, class RelTy>
1366void EhInputSection::split(ArrayRef<RelTy> rels) {
1367 ArrayRef<uint8_t> d = content();
1368 const char *msg = nullptr;
1369 unsigned relI = 0;
1370 while (!d.empty()) {
1371 if (d.size() < 4) {
1372 msg = "CIE/FDE too small";
1373 break;
1374 }
1375 uint64_t size = endian::read32<ELFT::Endianness>(d.data());
1376 if (size == 0) // ZERO terminator
1377 break;
1378 uint32_t id = endian::read32<ELFT::Endianness>(d.data() + 4);
1379 size += 4;
1380 if (LLVM_UNLIKELY(size > d.size())) {
1381 // If it is 0xFFFFFFFF, the next 8 bytes contain the size instead,
1382 // but we do not support that format yet.
1383 msg = size == UINT32_MAX + uint64_t(4)
1384 ? "CIE/FDE too large"
1385 : "CIE/FDE ends past the end of the section";
1386 break;
1387 }
1388
1389 // Find the first relocation that points to [off,off+size). Relocations
1390 // have been sorted by r_offset.
1391 const uint64_t off = d.data() - content().data();
1392 while (relI != rels.size() && rels[relI].r_offset < off)
1393 ++relI;
1394 unsigned firstRel = -1;
1395 if (relI != rels.size() && rels[relI].r_offset < off + size)
1396 firstRel = relI;
1397 (id == 0 ? cies : fdes).emplace_back(Args: off, Args: this, Args&: size, Args&: firstRel);
1398 d = d.slice(N: size);
1399 }
1400 if (msg)
1401 Err(ctx&: file->ctx) << "corrupted .eh_frame: " << msg << "\n>>> defined in "
1402 << getObjMsg(offset: d.data() - content().data());
1403}
1404
1405// Return the offset in an output section for a given input offset.
1406uint64_t EhInputSection::getParentOffset(uint64_t offset) const {
1407 auto it = partition_point(
1408 Range: fdes, P: [=](EhSectionPiece p) { return p.inputOff <= offset; });
1409 if (it == fdes.begin() || it[-1].inputOff + it[-1].size <= offset) {
1410 it = partition_point(
1411 Range: cies, P: [=](EhSectionPiece p) { return p.inputOff <= offset; });
1412 if (it == cies.begin()) // invalid piece
1413 return offset;
1414 }
1415 if (it[-1].outputOff == -1) // invalid piece
1416 return offset - it[-1].inputOff;
1417 return it[-1].outputOff + (offset - it[-1].inputOff);
1418}
1419
1420static size_t findNull(StringRef s, size_t entSize) {
1421 for (unsigned i = 0, n = s.size(); i != n; i += entSize) {
1422 const char *b = s.begin() + i;
1423 if (std::all_of(first: b, last: b + entSize, pred: [](char c) { return c == 0; }))
1424 return i;
1425 }
1426 llvm_unreachable("");
1427}
1428
1429// Split SHF_STRINGS section. Such section is a sequence of
1430// null-terminated strings.
1431void MergeInputSection::splitStrings(StringRef s, size_t entSize) {
1432 const bool live = !(flags & SHF_ALLOC) || !getCtx().arg.gcSections;
1433 const char *p = s.data(), *end = s.data() + s.size();
1434 if (!std::all_of(first: end - entSize, last: end, pred: [](char c) { return c == 0; })) {
1435 Err(ctx&: getCtx()) << this << ": string is not null terminated";
1436 pieces.emplace_back(Args&: entSize, Args: 0, Args: false);
1437 return;
1438 }
1439 if (entSize == 1) {
1440 // Optimize the common case.
1441 do {
1442 size_t size = strlen(s: p);
1443 pieces.emplace_back(Args: p - s.begin(), Args: xxh3_64bits(data: StringRef(p, size)), Args: live);
1444 p += size + 1;
1445 } while (p != end);
1446 } else {
1447 do {
1448 size_t size = findNull(s: StringRef(p, end - p), entSize);
1449 pieces.emplace_back(Args: p - s.begin(), Args: xxh3_64bits(data: StringRef(p, size)), Args: live);
1450 p += size + entSize;
1451 } while (p != end);
1452 }
1453}
1454
1455// Split non-SHF_STRINGS section. Such section is a sequence of
1456// fixed size records.
1457void MergeInputSection::splitNonStrings(ArrayRef<uint8_t> data,
1458 size_t entSize) {
1459 size_t size = data.size();
1460 assert((size % entSize) == 0);
1461 const bool live = !(flags & SHF_ALLOC) || !getCtx().arg.gcSections;
1462
1463 pieces.resize_for_overwrite(N: size / entSize);
1464 for (size_t i = 0, j = 0; i != size; i += entSize, j++)
1465 pieces[j] = {i, (uint32_t)xxh3_64bits(data: data.slice(N: i, M: entSize)), live};
1466}
1467
1468template <class ELFT>
1469MergeInputSection::MergeInputSection(ObjFile<ELFT> &f,
1470 const typename ELFT::Shdr &header,
1471 StringRef name)
1472 : InputSectionBase(f, header, name, InputSectionBase::Merge) {}
1473
1474MergeInputSection::MergeInputSection(Ctx &ctx, StringRef name, uint32_t type,
1475 uint64_t flags, uint64_t entsize,
1476 ArrayRef<uint8_t> data)
1477 : InputSectionBase(ctx.internalFile, name, type, flags, /*link=*/0,
1478 /*info=*/0,
1479 /*addralign=*/entsize, entsize, data,
1480 SectionBase::Merge) {}
1481
1482// This function is called after we obtain a complete list of input sections
1483// that need to be linked. This is responsible to split section contents
1484// into small chunks for further processing.
1485//
1486// Note that this function is called from parallelForEach. This must be
1487// thread-safe (i.e. no memory allocation from the pools).
1488void MergeInputSection::splitIntoPieces() {
1489 assert(pieces.empty());
1490
1491 if (flags & SHF_STRINGS)
1492 splitStrings(s: toStringRef(Input: contentMaybeDecompress()), entSize: entsize);
1493 else
1494 splitNonStrings(data: contentMaybeDecompress(), entSize: entsize);
1495}
1496
1497SectionPiece &MergeInputSection::getSectionPiece(uint64_t offset) {
1498 if (content().size() <= offset) {
1499 Err(ctx&: getCtx()) << this << ": offset is outside the section";
1500 return pieces[0];
1501 }
1502 return partition_point(
1503 Range&: pieces, P: [=](SectionPiece p) { return p.inputOff <= offset; })[-1];
1504}
1505
1506// Return the offset in an output section for a given input offset.
1507uint64_t MergeInputSection::getParentOffset(uint64_t offset) const {
1508 const SectionPiece &piece = getSectionPiece(offset);
1509 return piece.outputOff + (offset - piece.inputOff);
1510}
1511
1512template InputSection::InputSection(ObjFile<ELF32LE> &, const ELF32LE::Shdr &,
1513 StringRef);
1514template InputSection::InputSection(ObjFile<ELF32BE> &, const ELF32BE::Shdr &,
1515 StringRef);
1516template InputSection::InputSection(ObjFile<ELF64LE> &, const ELF64LE::Shdr &,
1517 StringRef);
1518template InputSection::InputSection(ObjFile<ELF64BE> &, const ELF64BE::Shdr &,
1519 StringRef);
1520
1521template void InputSection::writeTo<ELF32LE>(Ctx &, uint8_t *);
1522template void InputSection::writeTo<ELF32BE>(Ctx &, uint8_t *);
1523template void InputSection::writeTo<ELF64LE>(Ctx &, uint8_t *);
1524template void InputSection::writeTo<ELF64BE>(Ctx &, uint8_t *);
1525
1526template RelsOrRelas<ELF32LE>
1527InputSectionBase::relsOrRelas<ELF32LE>(bool) const;
1528template RelsOrRelas<ELF32BE>
1529InputSectionBase::relsOrRelas<ELF32BE>(bool) const;
1530template RelsOrRelas<ELF64LE>
1531InputSectionBase::relsOrRelas<ELF64LE>(bool) const;
1532template RelsOrRelas<ELF64BE>
1533InputSectionBase::relsOrRelas<ELF64BE>(bool) const;
1534
1535template MergeInputSection::MergeInputSection(ObjFile<ELF32LE> &,
1536 const ELF32LE::Shdr &, StringRef);
1537template MergeInputSection::MergeInputSection(ObjFile<ELF32BE> &,
1538 const ELF32BE::Shdr &, StringRef);
1539template MergeInputSection::MergeInputSection(ObjFile<ELF64LE> &,
1540 const ELF64LE::Shdr &, StringRef);
1541template MergeInputSection::MergeInputSection(ObjFile<ELF64BE> &,
1542 const ELF64BE::Shdr &, StringRef);
1543
1544template EhInputSection::EhInputSection(ObjFile<ELF32LE> &,
1545 const ELF32LE::Shdr &, StringRef);
1546template EhInputSection::EhInputSection(ObjFile<ELF32BE> &,
1547 const ELF32BE::Shdr &, StringRef);
1548template EhInputSection::EhInputSection(ObjFile<ELF64LE> &,
1549 const ELF64LE::Shdr &, StringRef);
1550template EhInputSection::EhInputSection(ObjFile<ELF64BE> &,
1551 const ELF64BE::Shdr &, StringRef);
1552
1553template void EhInputSection::split<ELF32LE>();
1554template void EhInputSection::split<ELF32BE>();
1555template void EhInputSection::split<ELF64LE>();
1556template void EhInputSection::split<ELF64BE>();
1557

Provided by KDAB

Privacy Policy
Improve your Profiling and Debugging skills
Find out more

source code of lld/ELF/InputSection.cpp