Page MenuHomePhabricator

Please use GitHub pull requests for new patches. Phabricator shutdown timeline

Differential D85579 Diff 284115 lld/ELF/SyntheticSections.cpp

Changeset View

Standalone View

View Options

lld/ELF/SyntheticSections.cpp

Show First 20 Lines • Show All 2,693 Lines • ▼ Show 20 Linesvoid GdbIndexSection::initOutputSize() {
// Add the constant pool size if exists. // Add the constant pool size if exists.
if (!symbols.empty()) { if (!symbols.empty()) {
GdbSymbol &sym = symbols.back(); GdbSymbol &sym = symbols.back();
size += sym.nameOff + sym.name.size() + 1; size += sym.nameOff + sym.name.size() + 1;
} }
} }
static std::vector<InputSection *> getDebugInfoSections() { static std::vector<InputSection *> getDebugInfoSections() {
// Add at most one .debug_info for an InputFile, Also see the comment in
// LLDDwarfObj<ELFT>::LLDDwarfObj. If we do lightweight parsing in the future,
// note that isec->data() may uncompress the full content, which should be
// parallelized.
std::vector<InputSection *> ret; std::vector<InputSection *> ret;
DenseSet<InputFile *> seen;
for (InputSectionBase *s : inputSections) for (InputSectionBase *s : inputSections)
if (InputSection *isec = dyn_cast<InputSection>(s)) if (InputSection *isec = dyn_cast<InputSection>(s))
if (isec->name == ".debug_info") if (isec->name == ".debug_info" && seen.insert(isec->file).second)
ret.push_back(isec); ret.push_back(isec);
return ret; return ret;
} }
static std::vector<GdbIndexSection::CuEntry> readCuList(DWARFContext &dwarf) { static std::vector<GdbIndexSection::CuEntry> readCuList(DWARFContext &dwarf) {
std::vector<GdbIndexSection::CuEntry> ret; std::vector<GdbIndexSection::CuEntry> ret;
for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units()) for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units())
ret.push_back({cu->getOffset(), cu->getLength() + 4}); ret.push_back({cu->getOffset(), cu->getLength() + 4});
▲ Show 20 Lines • Show All 141 Lines • ▼ Show 20 Lines for (GdbSymbol &sym : ret) {
off += sym.name.size() + 1; off += sym.name.size() + 1;
} }
return ret; return ret;
} }
// Returns a newly-created .gdb_index section. // Returns a newly-created .gdb_index section.
template <class ELFT> GdbIndexSection *GdbIndexSection::create() { template <class ELFT> GdbIndexSection *GdbIndexSection::create() {
std::vector<InputSection *> sections = getDebugInfoSections(); std::vector<InputSection *> sections = getDebugInfoSections();
grimarUnsubmitted
Done
ReplyInline Actions

It doesn't seem the getDebugInfoSections helper and its name makes much sence now with this change?

  1. It is not clear what it returns anymore: it might return either CU or TU .debug_infos.
  2. You are not really using the result, since dobj.getInfoSection() is called instead to get debug info sections.
  3. It is just needed now to find the number of files with .debug_info sections.

So I think it should probably just be inlined and/or rewritten to collect a file set mentioned probably?

grimar: It doesn't seem the `getDebugInfoSections` helper and its name makes much sence now with this…
// .debug_gnu_pub{names,types} are useless in executables. // .debug_gnu_pub{names,types} are useless in executables.
// They are present in input object files solely for creating // They are present in input object files solely for creating
// a .gdb_index. So we can remove them from the output. // a .gdb_index. So we can remove them from the output.
for (InputSectionBase *s : inputSections) for (InputSectionBase *s : inputSections)
if (s->name == ".debug_gnu_pubnames" || s->name == ".debug_gnu_pubtypes") if (s->name == ".debug_gnu_pubnames" || s->name == ".debug_gnu_pubtypes")
s->markDead(); s->markDead();
std::vector<GdbChunk> chunks(sections.size()); std::vector<GdbChunk> chunks(sections.size());
std::vector<std::vector<NameAttrEntry>> nameAttrs(sections.size()); std::vector<std::vector<NameAttrEntry>> nameAttrs(sections.size());
parallelForEachN(0, sections.size(), [&](size_t i) { parallelForEachN(0, sections.size(), [&](size_t i) {
// To keep memory usage low, we don't want to keep cached DWARFContext, so // To keep memory usage low, we don't want to keep cached DWARFContext, so
// avoid getDwarf() here. // avoid getDwarf() here.
ObjFile<ELFT> *file = sections[i]->getFile<ELFT>(); ObjFile<ELFT> *file = sections[i]->getFile<ELFT>();
DWARFContext dwarf(std::make_unique<LLDDwarfObj<ELFT>>(file)); DWARFContext dwarf(std::make_unique<LLDDwarfObj<ELFT>>(file));
auto &dobj = static_cast<const LLDDwarfObj<ELFT> &>(dwarf.getDWARFObj());
chunks[i].sec = sections[i]; // sections[i] may not be the compile unit. Use dobj.getInfoSection().
chunks[i].sec = dobj.getInfoSection();
chunks[i].compilationUnits = readCuList(dwarf); chunks[i].compilationUnits = readCuList(dwarf);
chunks[i].addressAreas = readAddressAreas(dwarf, sections[i]); chunks[i].addressAreas = readAddressAreas(dwarf, chunks[i].sec);
nameAttrs[i] = readPubNamesAndTypes<ELFT>( nameAttrs[i] = readPubNamesAndTypes<ELFT>(dobj, chunks[i].compilationUnits);
static_cast<const LLDDwarfObj<ELFT> &>(dwarf.getDWARFObj()),
chunks[i].compilationUnits);
}); });
grimarUnsubmitted
Done
ReplyInline Actions

Previously, parallelForEachN loop assumed that there is a one .debug_info section in an object.
Now it is not true and parallelizing for sections by itself (instead of doing it for files) doesn't look optimal.

Also, you do not need to build a list sections, because you are not using them (you call dobj.getInfoSection(); anyways), so
you only need a list of files to work with.

With that we can have a simpler code:

  SetVector<ObjFile<ELFT> *> files;
  for (InputSectionBase *s : inputSections) {
    InputSection *isec = dyn_cast<InputSection>(s);
    if (!isec)
      continue;
    // .debug_gnu_pub{names,types} are useless in executables.
    // They are present in input object files solely for creating
    // a .gdb_index. So we can remove them from the output.
    if (s->name == ".debug_gnu_pubnames" || s->name == ".debug_gnu_pubtypes")
      s->markDead();
    else if (isec->name == ".debug_info")
      files.insert(isec->getFile<ELFT>());
  }

  std::vector<GdbChunk> chunks(files.size());
  std::vector<std::vector<NameAttrEntry>> nameAttrs(files.size());

  parallelForEachN(0, files.size(), [&](size_t i) {
    // To keep memory usage low, we don't want to keep cached DWARFContext, so
    // avoid getDwarf() here.
    DWARFContext dwarf(std::make_unique<LLDDwarfObj<ELFT>>(files[i]));
    auto &dobj = static_cast<const LLDDwarfObj<ELFT> &>(dwarf.getDWARFObj());
...
grimar: Previously, `parallelForEachN` loop assumed that there is a one `.debug_info` section in an…
MaskRayAuthorUnsubmitted
Done
ReplyInline Actions

Forgot to push this comment:

Thanks! I'll make a small change: SetVector<InputFile *>, just so that we don't need to instantiate nearly identical code for 4 instantiations of ELFT.

MaskRay: Forgot to push this comment: Thanks! I'll make a small change: `SetVector<InputFile *>`, just…
auto *ret = make<GdbIndexSection>(); auto *ret = make<GdbIndexSection>();
ret->chunks = std::move(chunks); ret->chunks = std::move(chunks);
ret->symbols = createSymbols(nameAttrs, ret->chunks); ret->symbols = createSymbols(nameAttrs, ret->chunks);
ret->initOutputSize(); ret->initOutputSize();
return ret; return ret;
} }
▲ Show 20 Lines • Show All 974 Lines • Show Last 20 Lines