[Lldb-commits] [lldb] Improve type and namespace lookup using parent chain (PR #108907)

[Felipe de Azevedo Piovezan via lldb-commits]

================
@@ -374,25 +377,40 @@ void DebugNamesDWARFIndex::GetFullyQualifiedType(
   m_fallback.GetFullyQualifiedType(context, callback);
 }
 
+bool DebugNamesDWARFIndex::SameAsEntryATName(
+    llvm::StringRef query_name, const DebugNames::Entry &entry) const {
+  auto maybe_dieoffset = entry.getDIEUnitOffset();
+  if (!maybe_dieoffset)
+    return false;
+
+  // [Optimization] instead of parsing the entry from dwo file, we simply
+  // check if the query_name can point to an entry of the same DIE offset.
+  // This greatly reduced number of dwo file parsed and thus improved the
+  // performance.
+  for (const DebugNames::Entry &query_entry :
+       entry.getNameIndex()->equal_range(query_name)) {
+    auto query_dieoffset = query_entry.getDIEUnitOffset();
+    if (!query_dieoffset)
+      continue;
+
+    if (*query_dieoffset == *maybe_dieoffset) {
+      return true;
+    } else if (*query_dieoffset > *maybe_dieoffset) {
+      // The pool entries of the same name are sequentially cluttered together
+      // so if the query name from `query_name` is after the target entry, this
+      // is definitely not the correct one, we can stop searching.
+      return false;
+    }
+  }
+  return false;
+}
+
 bool DebugNamesDWARFIndex::SameParentChain(
     llvm::ArrayRef<llvm::StringRef> parent_names,
     llvm::ArrayRef<DebugNames::Entry> parent_entries) const {
-
   if (parent_entries.size() != parent_names.size())
     return false;
 
-  auto SameAsEntryATName = [this](llvm::StringRef name,
-                                  const DebugNames::Entry &entry) {
-    // Peek at the AT_name of `entry` and test equality to `name`.
-    auto maybe_dieoffset = entry.getDIEUnitOffset();
-    if (!maybe_dieoffset)
-      return false;
-    DWARFUnit *unit = GetNonSkeletonUnit(entry);
-    if (!unit)
-      return false;
-    return name == unit->PeekDIEName(unit->GetOffset() + *maybe_dieoffset);
----------------
felipepiovezan wrote:

> but the evaluation is very fast 293ms which I do not think it is slow enough to measure perf. How do you get 5~6 seconds from this?

Oh, my bad, to get to 5s you would have to revert the original parent chain patch. It is still interesting that you are getting 0.2s for that experiment when I was getting 1.4s (after the parent chain patches). I have some time today and will try to repro this out of curiosity.

But I am glad this doesn't seem to regress perf!

https://github.com/llvm/llvm-project/pull/108907