[colbymchenry]

Summary

C++ looks up the members of an anonymous union or struct as members of the class it is declared in ([class.union.anon]). codegraph reads a class's data members from its own lines, skipping those of the functions and classes nested in it. The index keeps an anonymous union as a nested <anonymous> node, so its lines were skipped too.

Every generated protobuf message keeps its fields in union { Impl_ _impl_; };, so calls through _impl_ were guessed by the method's name:

• this_._impl_.type_url_.Destroy() reached Arena::Destroy;
• _impl_._cached_size_.Get() reached FieldArenaRep::Get;
• _impl_._has_bits_.Clear() reached the message's own Clear.

fmt's value keeps its argument in a union of string, custom and the rest in the same way.

This is the first PR of follow-up 1 from #2628 (C++ member chains, step 4).

Why this one: re-verified on main 81fd007

A temporary, log-only diagnostic on main recorded every member-chain call and its outcome. For each call left to the name strategies, it also recorded the first hop that couldn't be typed. Its dumps were byte-identical to an uninstrumented baseline.

repo	chain calls	linked	linked to nothing	left to bare-name guessing
rocksdb	9,749	5,461	3,638	650
protobuf	2,121	477	280	1,364
leveldb	236	214	15	7
fmt	111	76	12	23

Most of protobuf's largest buckets were _impl_:

• 402 untyped bases, whose declaration the scan found inside the union;
• 16 bases found nowhere;
• X::_impl_ as most of the 265 members not found.

Fix

cppNestedSpans leaves out an anonymous union or struct that has no name and no declarator (isCppAnonymousMembers): its head is a bare union { or struct {, and }; follows its closing brace. Two cases keep their members to themselves:

• a union with a declarator (union { … } u;);
• a partial specialization. The index also keeps it as <anonymous>, but its head names its template.

Validation

Kit A/B against main 81fd007 (leveldb, fmt, protobuf, rocksdb, redis):

repo	removed	added	retargeted	same target, now typed
protobuf	2	54	349	231
fmt	0	0	1	1
rocksdb	0	0	2	0
leveldb, redis	identical

• Retargeted: from bare-name guesses to the member's class:
  • ArenaAlign::IsDefault, FieldArenaRep::Get, Arena::Destroy and Formatter::Set → ArenaStringPtr;
  • FieldArenaRep::Get → CachedSize::Get;
  • the message's own Clear → HasBits::Clear or RepeatedPtrField::Clear;
  • Any::GetCachedSize/MergeFrom → SourceContext through source_context_;
  • Any::ByteSizeLong/New → MessageLite through ptr.message_value.
• Removed: both are ptr.string_value->clear() on a std::string*, which had reached Map::clear.
• Added: typed RepeatedPtrField and UntypedMapBase calls through _impl_ and map_, and sub.GetCachedSize() on locals typed through _impl_.
• fmt and rocksdb: their retargets pick the overload their argument fits, now that a union member has a type:
  • map(value_.uint128_value) takes native_uint128, not native_int128;
  • Add(…, props.oldest_ancestor_time) takes the uint64_t overload;
  • AppendProperty(…, oldest_ancestor_time, …) takes the template, not the const std::string& one.
• Hand check: about 45 changed sites across every group were all right.

Tests. __tests__/cpp-anonymous-union-members.test.ts has 5 tests:

• an out-of-class inline method;
• an in-class method reading a union member declared below it;
• a this_._impl_.… chain through a reference parameter;
• fmt's shape;
• a control: a union with a declarator, a named nested struct and a partial specialization keep their members to themselves.

The first four fail on main, with each namesake placed where the bare-name guess lands on it. The control passes on both.

Mutations:

• Killed: dropping the nested-span exclusion, the declarator check or the head check.
• Survived: dropping the <anonymous> name check. The head check already covers named types; the name check stays as a cheap early exit.
• Dropped from the change: a matching edit to the caller's-declaration check survived too, and with it removed the kit A/B was byte-identical, so it isn't in this PR.

Cost: protobuf was indexed four times, alternating main and fix: 117 s and 130 s on main, 120 s and 129 s with the fix. That is the same within noise. leveldb and redis stay byte-identical.

Full suite (--maxWorkers=4 --minWorkers=1, kernel built, CODEGRAPH_KERNEL_EXPECT=1): 677 files and 8,115 tests passed, 6 files and 82 tests were skipped, and none failed.

CHANGELOG: one bullet under [Unreleased] → ### Fixes, placed next to the member-chain bullet. Re-index C++ projects after upgrading.

Left for follow-up 1's next PRs

From the same diagnosis, before this PR:

• pimpl Rep (rocksdb 103, leveldb 7): the member's type is looked up from the header, where Rep is only forward-declared. This is the next candidate.
• classes the index doesn't hold (protobuf): descriptor.pb.h is over the 1 MiB cap. That accounts for about 470 unplaced hops, 84 bases in its methods and 19 final types.
• assorted rocksdb shapes:
  • untyped bases, mostly unreadable auto and structured bindings (224);
  • members whose type spans several lines (TraceStats, 38);
  • iterator classes reached through operator-> (60), which belong to follow-up 2.

🤖 Generated with Claude Code