[lld] [lld][ELF] Implement merged .debug_names section. (PR #86508)

Alexander Yermolovich via llvm-commits llvm-commits at lists.llvm.org
Thu Mar 28 13:35:34 PDT 2024


================
@@ -2690,6 +2693,770 @@ static uint32_t computeGdbHash(StringRef s) {
   return h;
 }
 
+template <class ELFT>
+DebugNamesSection<ELFT>::DebugNamesSection()
+    : SyntheticSection(0, SHT_PROGBITS, 1, ".debug_names") {}
+
+template <class ELFT>
+template <class RelTy>
+void DebugNamesSection<ELFT>::getNameRelocsImpl(
+    InputSection *sec, ArrayRef<RelTy> rels,
+    DenseMap<uint32_t, uint32_t> &relocs) {
+  for (auto &rel : rels) {
+    Symbol &sym = sec->getFile<ELFT>()->getRelocTargetSym(rel);
+    relocs[rel.r_offset] = sym.getVA(getAddend<ELFT>(rel));
+  }
+}
+
+template <class ELFT>
+void DebugNamesSection<ELFT>::getNameRelocs(
+    InputSectionBase *base, DenseMap<uint32_t, uint32_t> &relocs) {
+  auto *sec = cast<InputSection>(base);
+  const RelsOrRelas<ELFT> rels = sec->template relsOrRelas<ELFT>();
+  if (rels.areRelocsRel())
+    getNameRelocsImpl(sec, rels.rels, relocs);
+  else
+    getNameRelocsImpl(sec, rels.relas, relocs);
+}
+
+template <class ELFT> void DebugNamesSection<ELFT>::writeTo(uint8_t *buf) {
+  SmallVector<uint32_t, 0> mergedCuOffsets;
+  SmallVector<uint32_t, 0> mergedTuOffsets;
+  DenseMap<uint32_t, uint32_t> strOffsets;
+  SmallVector<DenseMap<uint32_t, uint32_t>, 0> chunksRelocs;
+  chunksRelocs.reserve(numChunks);
+
+  for (size_t i = 0; i < numChunks; ++i) {
+    DebugNamesOutputChunk &chunk = outputChunks[i];
+    InputSectionBase *base = inputDebugNamesSections[i];
+    DenseMap<uint32_t, uint32_t> &relocs = chunksRelocs.emplace_back();
+    getNameRelocs(base, relocs);
+
+    // Update CuOffsets list with new data
+    for (uint32_t cuOffset : chunk.compilationUnits)
+      mergedCuOffsets.push_back(chunk.sec->outSecOff + cuOffset);
+
+    // TODO: Update TuOffsets list with new data
+  }
+
+  // Update the entries with the relocated string offsets.
+  for (auto &stringEntry : mergedEntries) {
+    uint32_t oldOffset = stringEntry.stringOffsetOffset;
+    uint32_t idx = stringEntry.chunkIdx;
+    stringEntry.relocatedEntryOffset = chunksRelocs[idx][oldOffset];
+  }
+
+  // Write out bytes for merged section.
+
+  // Write the header.
+  endian::write32<ELFT::TargetEndianness>(buf + 0, mergedHdr.UnitLength);
+  endian::write16<ELFT::TargetEndianness>(buf + 4, mergedHdr.Version);
+  endian::write32<ELFT::TargetEndianness>(buf + 8, mergedHdr.CompUnitCount);
+  endian::write32<ELFT::TargetEndianness>(buf + 12,
+                                          mergedHdr.LocalTypeUnitCount);
+  endian::write32<ELFT::TargetEndianness>(buf + 16,
+                                          mergedHdr.ForeignTypeUnitCount);
+  endian::write32<ELFT::TargetEndianness>(buf + 20, mergedHdr.BucketCount);
+  endian::write32<ELFT::TargetEndianness>(buf + 24, mergedHdr.NameCount);
+  endian::write32<ELFT::TargetEndianness>(buf + 28, mergedHdr.AbbrevTableSize);
+  endian::write32<ELFT::TargetEndianness>(buf + 32,
+                                          mergedHdr.AugmentationStringSize);
+  buf += 36;
+  memcpy(buf, mergedHdr.AugmentationString.c_str(), 8);
+  buf += 8;
+
+  // Write the CU list.
+  for (uint32_t offset : mergedCuOffsets) {
+    endian::write32<ELFT::TargetEndianness>(buf + 0, offset);
+    buf += 4;
+  }
+
+  // Write the local TU list.
+  // TODO: Fix this, once we get everything working without TUs.
+  if (mergedHdr.LocalTypeUnitCount != 0)
+    warn(".debug_names: type units are not handled in merged index");
+
+  // Write the foreign TU list.
+  // Currently LLVM does not emit foreign type units, so there should
+  // be nothing here to emit.
+  // TODO: Fix this, once we get everything working wtihout TUs.
+  if (mergedHdr.ForeignTypeUnitCount != 0)
+    warn(".debug_names: type units are not handled in merged index");
+
+  // Write the hash table.
+  // ... Write the buckets
+  uint32_t idx = 1;
+  for (const auto &bucket : bucketList) {
+    if (!bucket.empty())
+      endian::write32<ELFT::TargetEndianness>(buf + 0, idx);
+    idx += bucket.size();
+    buf += 4;
+  }
+
+  // ...Write the hashes
+  for (const auto &bucket : bucketList) {
+    for (const auto &entry : bucket) {
+      uint32_t hashValue = entry->hashValue;
+      endian::write32<ELFT::TargetEndianness>(buf + 0, hashValue);
+      buf += 4;
+    }
+  }
+
+  // Write the string offsets.
+  for (const auto &entry : mergedEntries) {
+    endian::write32<ELFT::TargetEndianness>(buf + 0,
+                                            entry.relocatedEntryOffset);
+    buf += 4;
+  }
+
+  // Write the entry offsets.
+  for (const auto &entry : mergedEntries) {
+    endian::write32<ELFT::TargetEndianness>(buf + 0, entry.entryOffset);
+    buf += 4;
+  }
+
+  // Write the abbrev table.
+  for (const auto *abbrev : mergedAbbrevTable) {
+    buf += encodeULEB128(abbrev->code, buf);
+    buf += encodeULEB128(abbrev->tag, buf);
+    for (auto attr : abbrev->attributes) {
+      buf += encodeULEB128(attr.Index, buf);
+      buf += encodeULEB128(attr.Form, buf);
+    }
+    endian::write16<ELFT::TargetEndianness>(buf + 0, 0); // attribute sentinels.
+    buf += 2;
+  }
+  *buf++ = 0; // abbrev table sentinel
+
+  // Write the entry pool.
+  for (const auto &stringEntry : mergedEntries) {
+    // Write all the entries for the string.
+    for (const auto &entry : stringEntry.indexEntries) {
+      buf += encodeULEB128(entry->abbrevCode, buf);
+      for (const auto &value : entry->attrValues) {
+        if (value.attrSize > 0) {
+          endian::write32<ELFT::TargetEndianness>(buf + 0, value.attrValue);
+          buf += value.attrSize;
+        }
+      }
+    }
+    *buf++ = 0; // Entry sentinel
+  }
+}
+
+template <class ELFT> bool DebugNamesSection<ELFT>::isNeeded() const {
+  return numChunks > 0;
+}
+
+template <class ELFT>
+static void readCompileUnitOffsets(
+    typename DebugNamesSection<ELFT>::DebugNamesSectionData &secData,
+    typename DebugNamesSection<ELFT>::DebugNamesOutputChunk &outputChunk,
+    DWARFDataExtractor &namesExtractor) {
+  uint64_t offset = secData.locs.CUsBase;
+  uint64_t *offsetPtr = &offset;
+  for (size_t i = 0; i < secData.hdr.CompUnitCount; ++i) {
+    Error err = Error::success();
+    // This call to namesExtractor can't fail; if there were a problem with the
+    // input, it would have been caught in the call to NamesIndex::extract, in
+    // DebugNamesSection::create.
+    uint32_t value = namesExtractor.getU32(offsetPtr, &err);
+    cantFail(std::move(err));
+    outputChunk.compilationUnits.push_back(value);
+  }
+}
+
+template <class ELFT>
+static void readEntryOffsets(
+    typename DebugNamesSection<ELFT>::DebugNamesSectionData &secData,
+    DWARFDataExtractor &namesExtractor) {
+  secData.entryOffsets = std::make_unique<uint32_t[]>(secData.hdr.NameCount);
+  uint64_t offset = secData.locs.EntryOffsetsBase;
+  uint64_t *offsetPtr = &offset;
+  for (size_t i = 0; i < secData.hdr.NameCount; ++i) {
+    Error err = Error::success();
+    // This call to namesExtractor can't fail; if there were a problem with the
+    // input, it would have been caught in the call to NamesIndex::extract, in
+    // DebugNamesSection::create.
+    uint32_t value = namesExtractor.getU32(offsetPtr, &err);
+    cantFail(std::move(err));
+    secData.entryOffsets.get()[i] = value;
+  }
+}
+
+template <class ELFT>
+static void readAttributeValues(
+    SmallVector<typename DebugNamesSection<ELFT>::AttrValueData, 3> &values,
+    typename DebugNamesSection<ELFT>::DebugNamesInputChunk &chunk,
+    uint64_t &offset,
+    typename DebugNamesSection<ELFT>::DebugNamesSectionData &secData,
+    int32_t &parentOffset, DWARFDataExtractor &namesExtractor,
+    const DWARFDebugNames::Abbrev &abbrev) {
+  const LLDDWARFSection &namesSection = *chunk.namesSection;
+  uint64_t *offsetPtr = &offset;
+  typename DebugNamesSection<ELFT>::AttrValueData cuOrTuAttr = {0, 0};
+  for (auto attr : abbrev.Attributes) {
+    Error err = Error::success();
+    typename DebugNamesSection<ELFT>::AttrValueData newAttr;
+    uint32_t value;
+    if (attr.Index == DW_IDX_parent && attr.Form != DW_FORM_flag_present &&
+        attr.Form != DW_FORM_ref4)
+      errorOrWarn(toString(namesSection.sec) +
+                  Twine(": invalid form for DW_IDX_parent"));
+    switch (attr.Form) {
+    case DW_FORM_flag_present: {
+      // Currently only DW_IDX_parent attributes (in .debug_names) can
+      // have this form. This form does not have a real value (nothing is
+      // emitted for it).
+      if (attr.Index != DW_IDX_parent)
+        errorOrWarn(toString(namesSection.sec) +
+                    Twine(": invalid form for attribute"));
+      break;
+    }
+    case DW_FORM_data1:
+    case DW_FORM_ref1: {
+      newAttr.attrValue = namesExtractor.getU8(offsetPtr, &err);
+      newAttr.attrSize = 1;
+      break;
+    }
+    case DW_FORM_data2:
+    case DW_FORM_ref2: {
+      value = namesExtractor.getU16(offsetPtr, &err);
+      newAttr.attrValue = value;
+      newAttr.attrSize = 2;
+      break;
+    }
+    case DW_FORM_data4:
+    case DW_FORM_ref4: {
+      value = namesExtractor.getU32(offsetPtr, &err);
+      newAttr.attrValue = value;
+      newAttr.attrSize = 4;
+      if (attr.Index == dwarf::DW_IDX_parent)
+        parentOffset = value + secData.locs.EntriesBase;
+      break;
+    }
+    case DW_FORM_data8:
+    case DW_FORM_ref8:
+    case DW_FORM_ref_sig8: {
+      value = namesExtractor.getU64(offsetPtr, &err);
+      newAttr.attrValue = value;
+      newAttr.attrSize = 8;
+      break;
+    }
+    default: {
+      errorOrWarn(toString(namesSection.sec) +
+                  Twine(": unrecognized form encoding ") + Twine(attr.Form) +
+                  Twine(" in .debug_names abbrev table"));
+      break;
+    }
+    }
+    if (err)
+      errorOrWarn(toString(namesSection.sec) +
+                  Twine(": error while reading attributes: ") +
+                  toString(std::move(err)));
+    if (attr.Index == DW_IDX_compile_unit || attr.Index == DW_IDX_type_unit)
+      // Save it to put it at the end of the attributes list.
+      cuOrTuAttr = newAttr;
+    else if (attr.Form != DW_FORM_flag_present)
+      values.push_back(newAttr);
+  }
+
+  // Make sure the CU or TU index attribute is the last one in the list.
+  values.push_back(cuOrTuAttr);
+}
+
+template <class ELFT>
+static void
+readEntry(typename DebugNamesSection<ELFT>::NamedEntry &stringEntry,
+          typename DebugNamesSection<ELFT>::DebugNamesInputChunk &chunk,
+          typename DebugNamesSection<ELFT>::DebugNamesSectionData &secData,
+          DWARFDataExtractor &namesExtractor,
+          const DWARFDebugNames::NameIndex &ni) {
+  std::unique_ptr<LLDDWARFSection> &namesSection = chunk.namesSection;
+  uint64_t offset = stringEntry.entryOffset;
+  // Each entry ends with a zero 'sentinel' value
+  while (offset < namesSection->Data.size() &&
+         namesSection->Data[offset] != 0) {
+    // Read & store all entries (for the same string)
+    auto entry =
+        std::make_unique<typename DebugNamesSection<ELFT>::IndexEntry>();
+    entry->poolOffset = offset;
+    Error err = Error::success();
+    // This call to namesExtractor can't fail; if there were a problem with the
+    // input, it would have been caught in the call to NamesIndex::extract, in
+    // DebugNamesSection::create.
+    uint64_t ulebValue = namesExtractor.getULEB128(&offset, &err);
+    cantFail(std::move(err));
+    entry->abbrevCode = static_cast<uint32_t>(ulebValue);
+    const auto &abbrevs = ni.getAbbrevs();
+    auto abbrevIt = abbrevs.find_as(entry->abbrevCode);
+    if (abbrevIt != abbrevs.end()) {
+      const DWARFDebugNames::Abbrev &abbrev = *abbrevIt;
+      readAttributeValues<ELFT>(entry->attrValues, chunk, offset, secData,
+                                entry->parentOffset, namesExtractor, abbrev);
+      stringEntry.indexEntries.push_back(std::move(entry));
+    } else
+      errorOrWarn(toString(chunk.namesSection->sec) +
+                  Twine(": invalid abbrev code in entry"));
+  }
+  if (offset >= namesSection->Data.size())
+    errorOrWarn(
+        toString(chunk.namesSection->sec) +
+        Twine(": encountered unexpected end of section while reading entry"));
+}
+
+template <class ELFT>
+static void
+readEntries(typename DebugNamesSection<ELFT>::DebugNamesSectionData &secData,
+            typename DebugNamesSection<ELFT>::DebugNamesInputChunk &chunk,
+            DWARFDataExtractor &namesExtractor, DataExtractor &strExtractor,
+            const DWARFDebugNames::NameIndex &ni) {
+  // Temporary map from entry offsets to address (pointer) of entry with that
+  // offset; used to find parent entries quickly.
+  DenseMap<uint32_t, typename DebugNamesSection<ELFT>::IndexEntry *> offsetMap;
+  // Reserve sizes for this chunk's hashes & namedEntries.
+  chunk.hashValues.reserve(secData.hdr.NameCount);
+  secData.namedEntries.reserve(secData.hdr.NameCount);
+  // Calculate the Entry Offsets, create initial records.
+  for (uint32_t i = 0, e = secData.hdr.NameCount; i != e; ++i) {
+    // Get string value
+    typename DebugNamesSection<ELFT>::NamedEntry stringEntry;
+    stringEntry.entryOffset =
+        secData.locs.EntriesBase + secData.entryOffsets[i];
+    uint64_t strOffsetOffset =
+        secData.locs.StringOffsetsBase + i * secData.dwarfSize;
+    stringEntry.stringOffsetOffset = strOffsetOffset;
+    uint64_t strOffset =
+        namesExtractor.getRelocatedValue(secData.dwarfSize, &strOffsetOffset);
+    StringRef name = strExtractor.getCStrRef(&strOffset);
+    stringEntry.name = name.data();
+    // Calculate hash
+    stringEntry.hashValue = caseFoldingDjbHash(name);
+    chunk.hashValues.push_back(stringEntry.hashValue);
+    // Read String Entry
+    readEntry<ELFT>(stringEntry, chunk, secData, namesExtractor, ni);
+    // Add index entries & offsets to our temporary map
+    for (const auto &entry : stringEntry.indexEntries)
+      offsetMap[entry->poolOffset] = entry.get();
+    secData.namedEntries.push_back(std::move(stringEntry));
+  }
+  // Find/assign pointers to any 'real' parent entries (needed to find correct
+  // parent entry offsets in merged data).
+  for (auto &stringEntry : secData.namedEntries)
+    for (auto &entry : stringEntry.indexEntries)
+      if (entry->parentOffset != -1)
+        entry->parentEntry = offsetMap[entry->parentOffset];
----------------
ayermolo wrote:

How will this handle when 
module 0
Entry0 0xOffset1 {...}
....
EntryX {
...
ParrentOffst 0xOffset1
}
module 1
Entry0 0xOffset1 {...}
....
EntryX {
...
ParrentOffst 0xOffset1
}

https://github.com/llvm/llvm-project/pull/86508


More information about the llvm-commits mailing list