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

[Fangrui Song via llvm-commits]

================
@@ -2690,6 +2693,759 @@ static uint32_t computeGdbHash(StringRef s) {
   return h;
 }
 
+DebugNamesSection::DebugNamesSection()
+    : SyntheticSection(0, SHT_PROGBITS, 1, ".debug_names") {}
+
+template <class ELFT, class RelTy>
+void DebugNamesSection::getNameRelocsImpl(
+    InputSection *sec, ArrayRef<RelTy> rels,
+    llvm::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::getNameRelocs(
+    InputSectionBase *base, llvm::DenseMap<uint32_t, uint32_t> &relocs) {
+  auto *sec = cast<InputSection>(base);
+  const RelsOrRelas<ELFT> rels = sec->template relsOrRelas<ELFT>();
+  if (rels.areRelocsRel())
+    getNameRelocsImpl<ELFT>(sec, rels.rels, relocs);
+  else
+    getNameRelocsImpl<ELFT>(sec, rels.relas, relocs);
+}
+
+void DebugNamesSection::writeTo(uint8_t *buf) { invokeELFT(writeToImpl, buf); }
+
+template <class ELFT> void DebugNamesSection::writeToImpl(uint8_t *buf) {
+  SmallVector<uint32_t, 0> mergedCuOffsets;
+  SmallVector<uint32_t, 0> mergedTuOffsets;
+  llvm::DenseMap<uint32_t, uint32_t> strOffsets;
+  SmallVector<llvm::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];
+    llvm::DenseMap<uint32_t, uint32_t> relocs;
+    getNameRelocs<ELFT>(base, relocs);
+    chunksRelocs.push_back(std::move(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) {
+    size_t uleb_size = encodeULEB128(abbrev->code, buf);
+    buf += uleb_size;
+    uleb_size = encodeULEB128(abbrev->tag, buf);
+    buf += uleb_size;
+    for (auto attr : abbrev->attributes) {
+      uleb_size = encodeULEB128(attr.Index, buf);
+      buf += uleb_size;
+      uleb_size = encodeULEB128(attr.Form, buf);
+      buf += uleb_size;
+    }
+    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.
+    size_t uleb_size;
+    for (const auto &entry : stringEntry.indexEntries) {
+      uleb_size = encodeULEB128(entry->abbrevCode, buf);
+      buf += uleb_size;
+      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
+  }
+}
+
+bool DebugNamesSection::isNeeded() const { return numChunks > 0; }
+
+template <class ELFT>
+static void
+readCompileUnitOffsets(struct DebugNamesSection::DebugNamesSectionData &secData,
+                       DebugNamesSection::DebugNamesInputChunk &inputChunk,
+                       DebugNamesSection::DebugNamesOutputChunk &outputChunk,
+                       llvm::DWARFDataExtractor &namesExtractor) {
+  uint64_t offset = secData.locs.CUsBase;
+  uint64_t *offsetPtr = &offset;
+  for (size_t i = 0; i < secData.hdr.CompUnitCount; ++i) {
+    llvm::Error err = llvm::Error::success();
+    uint32_t value = namesExtractor.getU32(offsetPtr, &err);
----------------
MaskRay wrote:

repeated bound checking using DataExtractor is inefficient.

`getU32` has an overload with a `count` parameter to read multiple values.

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