fix(cpp): a declaration with a qualifier or annotation around its name gives its type to calls on it

[colbymchenry]

Summary

C++ receiver inference reads a declaration as a type followed by the name, and took whatever stood right before the name for the type:

• Iterator* const iter = iter_guard.get(); (facebook/rocksdb) read as the type const, which names nothing. The scan then went on to an earlier declaration of iter, or the call was guessed: iter->SeekToFirst() reached the unrelated ArenaWrappedDBIter at 0.65.
• Slice const& key, a type written before its const, did the same.
• io::EpsCopyOutputStream* PROTOBUF_NONNULL stream, one of protobuf's ~3,600 nullability annotations, read as a type named PROTOBUF_NONNULL, so calls on stream got no edge.
• MemTable* imm_ GUARDED_BY(mutex_);, a leveldb thread-safety annotation, wasn't read at all: the name has to be followed by ;, =, ,, ) and so on. Calls on imm_ were guessed.

What changed (name-matcher.ts)

• Reading past what surrounds the name. buildDeclaratorRegex, and CPP_DECLARED_TYPE_TAIL for range-for, read past:
  • cv-qualifiers and __restrict before or after a * or & (T* const, T const&, T* const*);
  • nullability annotations after a * or &: abseil's and clang's (the words in CPP_NULLABILITY), and ALL-CAPS macros with an underscore, the rule c-cpp.ts's blankCppPointerAnnotationMacros uses for the parser;
  • an ALL-CAPS annotation macro's call after the name (GUARDED_BY(mutex_)).
• The type without its annotations. matchCppDeclarator hands the readers the type without them (cppDeclaratorType): Arena* PROTOBUF_NULLABLE becomes Arena*.
  • cv-qualifiers stay, since normalizeCppTypeName, the pointer check, cppClassWritten and the smart-pointer helpers already read past them.
  • An ALL-CAPS word is an annotation only after a * or &, so a type's own name stays (OS_HANDLE* const h).
  • Template arguments are left as written.
• Readers that use it:
  • the call's own file and its same-named header (inferCppReceiverType);
  • class members (cppDataMemberIn, fix(cpp): a call on a data member of the caller's class reaches its type's method, wherever the class is declared #2533);
  • range-for variables (cppRangeForDeclaration);
  • fix(cpp): a call a method makes with no receiver links to its class's own or inherited method #2479's local-shadowing check (isCppLocalName), through the regex alone: std::function<void()> const Finish now shadows a method Finish.

Validation

Site-keyed edge diffs (diff-sites.mjs), main df3ce972 vs this branch, both built from the same tree, kernel included:

repo	removed	added	retargeted	metadata-only
google/leveldb	18	52	60	81
fmtlib/fmt	3	13	6	60
facebook/rocksdb	52	232	141	402
protocolbuffers/protobuf	17	316	6	26
redis/redis (C, control)	0	0	0	0

Nodes and files are identical between the arms.

Removed (90), every site read. 83 were wrong guesses whose receivers now have a known type:

• Library types:
  • leveldb's GUARDED_BY containers had reached Slice::begin / end / clear (17): writers_.begin() on a std::deque<Writer*>, file_map_.end() on a std::map typedef.
  • rocksdb blob_files.size() had reached CompactionInputFiles::size (27).
  • rocksdb's other library-type calls (std::map, std::string, std::shared_ptr) had reached toku::omt::find, gtest's FilePath and others (22).
• protobuf descriptor classes whose accessors the index doesn't hold (fix(cpp): a call on a declared pointer, value or reference reaches the class its type names, not a namesake #2493's policy, 17):
  • descriptor_->full_name() on const Descriptor* const descriptor_ had reached EnumValueDescriptor::full_name;
  • a test's file_descriptor->options() had reached upb::FileDefPtr::options.

The other 7 lose a guess that happened to be right. In each, the declared type's class is one the index can't follow, and a declaration of that type without the qualifier gets no edge on main either:

• 6 where the class is indexed <anonymous>:
  • out-of-line nested definitions: leveldb SkipList<K,C>::Node (Node* const head_), rocksdb BlockBasedTable::Rep (const Rep* const rep, 2);
  • a partial specialization: fmt's vendored gmock FunctionMocker<R(Args...)> (FunctionMocker<F>* const, 3).
• 1 where the type is a template parameter named like a project class. rocksdb's InlineSkipList<Comparator> member Comparator const compare_ is looked up as rocksdb::Comparator, which has no decode_key. const Comparator compare_ gets the same on main.

Retargeted (213), every group read, all wrong→right. Calls on * const members and locals now reach the declared class instead of a namesake:

• rocksdb:
  • iter->SeekToFirst() after Iterator* const iter = iter_guard.get() (30, ArenaWrappedDBIter → IteratorBase);
  • Env* const env_ (50, CompositeEnv / PosixEnv / SpecialEnv → Env);
  • const SliceTransform* const prefix_extractor_ (17);
  • InstrumentedMutex* const mutex_ (4: was port::Mutex at 0.9);
  • SpecialEnv* const env → EnvWrapper::CreateDirIfMissing, DBImpl* const db_ in a local struct, and the rest.
• leveldb:
  • Env* const env_ (34, InMemoryEnv → Env);
  • DBIter's Iterator* const iter_ (21: was DBIter's own methods);
  • const Comparator* const user_comparator_, Reporter* const reporter_.
• protobuf: MessageFactory* const message_factory_, io::ZeroCopyOutputStream* const output_.
• fmt's vendored gtest: UnitTestImpl* const unit_test_, ExpectationBase* const locals.

Added (613): calls on the receivers now read, and chains through them.

• protobuf:
  • 176 on io::EpsCopyOutputStream* PROTOBUF_NONNULL stream in generated _InternalSerialize;
  • 95 on MessageLite, plus Arena* PROTOBUF_NULLABLE arena, io::Printer* PROTOBUF_NONNULL p and others;
  • 3 at 0.65 / 0.7 by the name strategies, all right, which the type misread as PROTOBUF_NONNULL / PROTOBUF_NULLABLE had blocked. Two are on const AccessorType* PROTOBUF_NONNULL accessor_, a template-dependent alias, which goes to the name strategies as such declarations do. One is on Chunk* PROTOBUF_NULLABLE head_, whose struct is defined out of line.
• rocksdb:
  • ColumnFamilyData* const cfd = … and chains through it (versions->GetColumnFamilySet()->GetDefault()), VersionStorageInfo* const storage_info, const InternalKeyComparator* const icmp_ and more;
  • 3 are wrong: Node* const head_ in InlineSkipList reaches a Node in hash_linklist_rep.cc's anonymous namespace. Main has the same lookup bug for 29 plain Node* locals in those headers (x->Next(…)).
• leveldb: the GUARDED_BY members (imm_, bg_error_, versions_, snapshots_) and the * const ones.

Metadata-only (569): the same target, now at 0.9.

Tests: __tests__/cpp-qualified-declarator-receiver.test.ts, 10 cases:

• 9 fail on main:
  • Iterator* const, with the nested namesake beside it, which main reached at 0.9;
  • Slice const& / Slice const*;
  • PROTOBUF_NULLABLE / absl_nonnull parameters, locals and members;
  • GUARDED_BY members;
  • auto* const, smart-pointer and range-for variables, one with an annotation;
  • an ALL-CAPS type name;
  • . and -> on qualified declarations;
  • header globals;
  • a std::function<void()> const parameter that shadows a method.
• 1 guards a declaration without qualifiers.
• 11 mutations, one per guard or branch; 10 fail a test. The survivor drops cppDeclaratorType's template-argument check, which keeps an argument like Foo<&GLOBAL_VAR> as written; no code in these repos depends on it.
• fix(cpp): a call on a data member of the caller's class reaches its type's method, wherever the class is declared #2533's member test still fails without either of its own two guards: a declarator inside parentheses, and a name an argument list follows.
• Full suite: 6,981 passed, 0 failed, 82 skipped. fix(cpp): a call on a data member of the caller's class reaches its type's method, wherever the class is declared #2533's test that a method's parameter is no member used T* const locals the scan couldn't read; it now declares none, since a local the scan reads decides first.

Together with #2537

#2537, bare calls in gtest bodies reaching the fixture's methods, is independent. The two branches touch different parts of name-matcher.ts and conflict only in CHANGELOG: keep both bullets. Both applied to df3ce972 pass the 28 C/C++ test files (335 tests).

Not in this PR

• Classes the index names <anonymous>: out-of-line definitions of nested classes (struct SkipList<K,C>::Node {…}, struct BlockBasedTable::Rep {…}) and partial specializations. A declared receiver of such a type gets no edge, on main too.
• C++ lookup finds a class in another file's anonymous namespace: hash_linklist_rep.cc's Node for inlineskiplist.h's own Node* receivers, 29 edges on main.

🤖 Generated with Claude Code