[lld] [lld][ELF] Add --debug-names to create merged .debug_names. (PR #86508)

Fangrui Song via llvm-commits llvm-commits at lists.llvm.org
Thu Apr 18 12:44:04 PDT 2024


================
@@ -2711,6 +2714,609 @@ static uint32_t computeGdbHash(StringRef s) {
   return h;
 }
 
+// 4-byte alignment ensures that values in the hash lookup table and the name
+// table are aligned.
+DebugNamesBaseSection::DebugNamesBaseSection()
+    : SyntheticSection(0, SHT_PROGBITS, 4, ".debug_names") {}
+
+// Get the size of the .debug_names section header in bytes for DWARF32:
+static uint32_t getDebugNamesHeaderSize(uint32_t augmentationStringSize) {
+  return /* unit length */ 4 +
+         /* version */ 2 +
+         /* padding */ 2 +
+         /* CU count */ 4 +
+         /* TU count */ 4 +
+         /* Foreign TU count */ 4 +
+         /* Bucket Count */ 4 +
+         /* Name Count */ 4 +
+         /* Abbrev table size */ 4 +
+         /* Augmentation string size */ 4 +
+         /* Augmentation string */ augmentationStringSize;
+}
+
+static Expected<DebugNamesBaseSection::IndexEntry *>
+readEntry(uint64_t &offset, const DWARFDebugNames::NameIndex &ni,
+          uint64_t entriesBase, DWARFDataExtractor &namesExtractor,
+          const LLDDWARFSection &namesSec) {
+  auto ie = makeThreadLocal<DebugNamesBaseSection::IndexEntry>();
+  ie->poolOffset = offset;
+  Error err = Error::success();
+  uint64_t ulebVal = namesExtractor.getULEB128(&offset, &err);
+  if (err)
+    return createStringError(inconvertibleErrorCode(),
+                             "invalid abbrev code in entry: %s",
+                             toString(std::move(err)).c_str());
+  if (ulebVal <= UINT32_MAX)
+    ie->abbrevCode = static_cast<uint32_t>(ulebVal);
+  else
+    return createStringError(inconvertibleErrorCode(),
+                             "abbrev code in entry too large for DWARF32: %d",
+                             ulebVal);
+  auto it = ni.getAbbrevs().find_as(ie->abbrevCode);
+  if (it == ni.getAbbrevs().end())
+    return createStringError(inconvertibleErrorCode(),
+                             "entry abbrev code not found in abbrev table: %d",
+                             ie->abbrevCode);
+
+  DebugNamesBaseSection::AttrValue attr, cuAttr = {0, 0};
+  for (DWARFDebugNames::AttributeEncoding a : it->Attributes) {
+    if (a.Index == dwarf::DW_IDX_parent) {
+      if (a.Form == dwarf::DW_FORM_ref4) {
+        attr.attrValue = namesExtractor.getU32(&offset, &err);
+        attr.attrSize = 4;
+        ie->parentOffset = entriesBase + attr.attrValue;
+      } else if (a.Form != DW_FORM_flag_present)
+        return createStringError(inconvertibleErrorCode(),
+                                 "invalid form for DW_IDX_parent");
+    } else {
+      switch (a.Form) {
+      case DW_FORM_data1:
+      case DW_FORM_ref1: {
+        attr.attrValue = namesExtractor.getU8(&offset, &err);
+        attr.attrSize = 1;
+        break;
+      }
+      case DW_FORM_data2:
+      case DW_FORM_ref2: {
+        attr.attrValue = namesExtractor.getU16(&offset, &err);
+        attr.attrSize = 2;
+        break;
+      }
+      case DW_FORM_data4:
+      case DW_FORM_ref4: {
+        attr.attrValue = namesExtractor.getU32(&offset, &err);
+        attr.attrSize = 4;
+        break;
+      }
+      default:
+        return createStringError(
+            inconvertibleErrorCode(),
+            "unrecognized form encoding %d in abbrev table", a.Form);
+      }
+    }
+    if (err)
+      return createStringError(inconvertibleErrorCode(),
+                               "error while reading attributes: %s",
+                               toString(std::move(err)).c_str());
+    if (a.Index == DW_IDX_compile_unit)
+      cuAttr = attr;
+    else if (a.Form != DW_FORM_flag_present)
+      ie->attrValues.push_back(attr);
+  }
+
+  // Canonicalize abbrev by placing the CU/TU index at the end.
+  ie->attrValues.push_back(cuAttr);
+
+  return ie;
+}
+
+void DebugNamesBaseSection::parseDebugNames(
+    InputChunk &inputChunk, OutputChunk &chunk,
+    DWARFDataExtractor &namesExtractor, DataExtractor &strExtractor,
+    function_ref<SmallVector<uint32_t, 0>(
+        uint32_t numCus, const DWARFDebugNames::Header &,
+        const DWARFDebugNames::DWARFDebugNamesOffsets &)>
+        readOffsets) {
+  const LLDDWARFSection &namesSec = inputChunk.section;
+  DenseMap<uint32_t, IndexEntry *> offsetMap;
+  // Number of CUs seen in previous NameIndex sections within current chunk.
+  uint32_t numCus = 0;
+  for (const DWARFDebugNames::NameIndex &ni : *inputChunk.llvmDebugNames) {
+    NameData &nd = inputChunk.nameData.emplace_back();
+    nd.hdr = ni.getHeader();
+    if (nd.hdr.Format != DwarfFormat::DWARF32) {
+      errorOrWarn(toString(namesSec.sec) +
+                  Twine(": found DWARF64, which is currently unsupported"));
+      return;
+    }
+    if (nd.hdr.Version != 5) {
+      errorOrWarn(toString(namesSec.sec) + Twine(": unsupported version: ") +
+                  Twine(nd.hdr.Version));
+      return;
+    }
+    uint32_t dwarfSize = dwarf::getDwarfOffsetByteSize(DwarfFormat::DWARF32);
+    DWARFDebugNames::DWARFDebugNamesOffsets locs = ni.getOffsets();
+    if (locs.EntriesBase > namesExtractor.getData().size()) {
+      errorOrWarn(toString(namesSec.sec) +
+                  Twine(": entry pool start is beyond end of section"));
+      return;
+    }
+
+    SmallVector<uint32_t, 0> entryOffsets = readOffsets(numCus, nd.hdr, locs);
+
+    // Read the entry pool.
+    offsetMap.clear();
+    nd.nameEntries.resize(nd.hdr.NameCount);
+    for (auto i : seq(nd.hdr.NameCount)) {
+      NameEntry &ne = nd.nameEntries[i];
+      uint64_t strOffset = locs.StringOffsetsBase + i * dwarfSize;
+      ne.stringOffset = strOffset;
+      uint64_t strp = namesExtractor.getRelocatedValue(dwarfSize, &strOffset);
+      StringRef name = strExtractor.getCStrRef(&strp);
+      ne.name = name.data();
+      ne.hashValue = caseFoldingDjbHash(name);
+
+      // Read a series of index entries that end with abbreviation code 0.
+      uint64_t offset = locs.EntriesBase + entryOffsets[i];
+      while (offset < namesSec.Data.size() && namesSec.Data[offset] != 0) {
+        // Read & store all entries (for the same string).
+        Expected<IndexEntry *> ieOrErr =
+            readEntry(offset, ni, locs.EntriesBase, namesExtractor, namesSec);
+        if (!ieOrErr) {
+          errorOrWarn(toString(namesSec.sec) + ": " +
+                      toString(ieOrErr.takeError()));
+          return;
+        }
+        ne.indexEntries.push_back(std::move(*ieOrErr));
+      }
+      if (offset >= namesSec.Data.size())
+        errorOrWarn(toString(namesSec.sec) +
+                    Twine(": index entry is out of bounds"));
+
+      for (IndexEntry &ie : ne.entries())
+        offsetMap[ie.poolOffset] = &ie;
+    }
+
+    // Assign parent pointers, which will be used to update DW_IDX_parent index
+    // attributes. Note: offsetMap[0] does not exist, so parentOffset == 0 will
+    // get parentEntry == null as well.
+    for (NameEntry &ne : nd.nameEntries)
+      for (IndexEntry &ie : ne.entries())
+        ie.parentEntry = offsetMap.lookup(ie.parentOffset);
+    numCus += nd.hdr.CompUnitCount;
+  }
+}
+
+// Compute the form for output DW_IDX_compile_unit attributes, similar to
+// DIEInteger::BestForm. The input form (often DW_FORM_data1) may not hold all
+// the merged CU indices.
+std::pair<uint8_t, dwarf::Form> static getMergedCuCountForm(
+    uint32_t compUnitCount) {
+  if (compUnitCount > UINT16_MAX)
+    return {4, DW_FORM_data4};
+  if (compUnitCount > UINT8_MAX)
+    return {2, DW_FORM_data2};
+  return {1, DW_FORM_data1};
+}
+
+void DebugNamesBaseSection::computeHdrAndAbbrevTable(
+    MutableArrayRef<InputChunk> inputChunks) {
+  TimeTraceScope timeScope("Merge .debug_names", "hdr and abbrev table");
+  size_t numCu = 0;
+  hdr.Format = DwarfFormat::DWARF32;
+  hdr.Version = 5;
+  hdr.CompUnitCount = 0;
+  hdr.LocalTypeUnitCount = 0;
+  hdr.ForeignTypeUnitCount = 0;
+  hdr.AugmentationStringSize = 0;
+
+  // Compute CU and TU counts.
+  for (auto i : seq(numChunks)) {
+    InputChunk &inputChunk = inputChunks[i];
+    inputChunk.baseCuIdx = numCu;
+    numCu += chunks[i].compUnits.size();
+    for (const NameData &nd : inputChunk.nameData) {
+      hdr.CompUnitCount += nd.hdr.CompUnitCount;
+      // TODO: We don't handle type units yet, so LocalTypeUnitCount &
+      // ForeignTypeUnitCount are left as 0.
+      if (nd.hdr.LocalTypeUnitCount || nd. hdr.ForeignTypeUnitCount)
+        warn(toString(inputChunk.section.sec) +
+             Twine(": type units are not implemented"));
+      // If augmentation strings are not identical, use an empty string.
+      if (i == 0) {
+        hdr.AugmentationStringSize = nd.hdr.AugmentationStringSize;
+        hdr.AugmentationString = nd.hdr.AugmentationString;
+      } else if (hdr.AugmentationString != nd.hdr.AugmentationString) {
+        // There are conflicting augmentation strings, so it's best for the
+        // merged index to not use an augmentation string.
+        hdr.AugmentationStringSize = 0;
+        hdr.AugmentationString.clear();
+      }
+    }
+  }
+
+  // Create the merged abbrev table, uniquifyinng the input abbrev tables and
+  // computing mapping from old (per-cu) abbrev codes to new (merged) abbrev
+  // codes.
+  FoldingSet<Abbrev> abbrevSet;
+  // Determine the form for the DW_IDX_compile_unit attributes in the merged
+  // index. The input form may not be big enough for all CU indices.
+  dwarf::Form cuAttrForm = getMergedCuCountForm(hdr.CompUnitCount).second;
+  for (InputChunk &inputChunk : inputChunks) {
+    for (auto [i, ni] : enumerate(*inputChunk.llvmDebugNames)) {
+      for (const DWARFDebugNames::Abbrev &oldAbbrev : ni.getAbbrevs()) {
+        // Canonicalize abbrev by placing the CU/TU index at the end,
+        // similar to 'parseDebugNames'.
+        Abbrev abbrev;
+        DWARFDebugNames::AttributeEncoding cuAttr(DW_IDX_compile_unit,
+                                                  cuAttrForm);
+        abbrev.code = oldAbbrev.Code;
+        abbrev.tag = oldAbbrev.Tag;
+        for (DWARFDebugNames::AttributeEncoding a : oldAbbrev.Attributes) {
+          if (a.Index == DW_IDX_compile_unit)
+            cuAttr.Index = a.Index;
+          else
+            abbrev.attributes.push_back({a.Index, a.Form});
+        }
+        // Put the CU/TU index at the end of the attributes list.
+        abbrev.attributes.push_back(cuAttr);
+
+        // Profile the abbrev, get or assign a new code, then record the abbrev
+        // code mapping.
+        FoldingSetNodeID id;
+        abbrev.Profile(id);
+        uint32_t newCode;
+        void *insertPos;
+        if (Abbrev *existing = abbrevSet.FindNodeOrInsertPos(id, insertPos)) {
+          // Found it; we've already seen an identical abbreviation.
+          newCode = existing->code;
+        } else {
+          Abbrev *abbrev2 =
+              new (abbrevAlloc.Allocate()) Abbrev(std::move(abbrev));
+          abbrevSet.InsertNode(abbrev2, insertPos);
+          abbrevTable.push_back(abbrev2);
+          newCode = abbrevTable.size();
+          abbrev2->code = newCode;
+        }
+        inputChunk.nameData[i].abbrevCodeMap[oldAbbrev.Code] = newCode;
+      }
+    }
+  }
+
+  // Compute the merged abbrev table.
+  raw_svector_ostream os(abbrevTableBuf);
+  for (Abbrev *abbrev : abbrevTable) {
+    encodeULEB128(abbrev->code, os);
+    encodeULEB128(abbrev->tag, os);
+    for (DWARFDebugNames::AttributeEncoding a : abbrev->attributes) {
+      encodeULEB128(a.Index, os);
+      encodeULEB128(a.Form, os);
+    }
+    os.write("\0", 2); // attribute specification end
+  }
+  os.write(0); // abbrev table end
+  hdr.AbbrevTableSize = abbrevTableBuf.size();
+}
+
+void DebugNamesBaseSection::Abbrev::Profile(FoldingSetNodeID &id) const {
+  id.AddInteger(tag);
+  for (const DWARFDebugNames::AttributeEncoding &attr : attributes) {
+    id.AddInteger(attr.Index);
+    id.AddInteger(attr.Form);
+  }
+}
+
+std::pair<uint32_t, uint32_t> DebugNamesBaseSection::computeEntryPool(
+    MutableArrayRef<InputChunk> inputChunks) {
+  TimeTraceScope timeScope("Merge .debug_names", "entry pool");
+  // Collect and de-duplicate all the names (preserving all the entries).
+  // Speed it up using multithreading, as the number of symbols can be in the
+  // order of millions.
+  const size_t concurrency =
+      bit_floor(std::min<size_t>(config->threadCount, numShards));
+  const size_t shift = 32 - countr_zero(numShards);
+  uint8_t cuAttrSize = getMergedCuCountForm(hdr.CompUnitCount).first;
----------------
MaskRay wrote:

`const uint8_t cuAttrSize`

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


More information about the llvm-commits mailing list