fix(cpp): a call on an element read by subscript, statuses[i].IsNotFound(), calls the element's method

[colbymchenry]

Summary

This is the follow-up left after #2601: calls on an element read by subscript.

The extractor can't spell a subscripted receiver, so it records statuses[i].IsNotFound(), buffers[key].Flush() and iters[i]->Valid() by their bare names. The name strategies then gave those calls to whichever class had a method of that name. Examples:

• rocksdb's values[i].c_str() and column_family_names_[i].c_str() on std::strings went to googletest's FilePath::c_str at 0.9;
• rocksdb's first_key_result[0].columns() on PinnableAttributeGroups went to ArenaWrappedDBIter::columns;
• protobuf's message_generators_[i]->GenerateInterface(printer) went to ImmutableMessageGenerator rather than the MessageGenerator the vector holds.

Separately, a local an element initializes (auto& s = statuses[i], const auto& key = keys_[ki]) had no type.

What changed:

• Subscript typing: C++ expression typing (cppSubscriptElement) now reads the element a subscript gives:

  • an array's element;
  • what a pointer points at;
  • a library map's mapped value;
  • a sequence's element (library containers, RepeatedField / RepeatedPtrField / autovector);
  • what a class's operator[] returns.

  It gives nothing for a std::string's char, or for a class whose operator[] the index doesn't hold.

• Arrays: declarations now record their array dimensions, which the type text doesn't show: LRUCache shard_[kNumShards], std::vector<FileMetaData*> files_[kNumLevels], Status grid[2][3].

  • This covers locals, parameters and data members, and survives following an alias.
  • A subscript, *arr, arr-> and a range-for over an array read the type the array is declared with.

• More receivers read from the source: fix(cpp): a call on the member of the pair a map's iterator points at, it->second.Flush(), calls the value's method #2601's bare-name receiver reader (cppUnspelledReceiver) now takes any number of subscripts after a name or (*name).

  • Examples: table[0][1].Flush(), (*m)[key].Flush(), cfs[i].first.Flush().
  • It reads the line with its literals blanked, so a key like buffers["a]b"] doesn't end the subscript.

• Overload arguments: cppArgumentType drops its own one-level name[i] reader for the shared subscript typing. A map's values[key] passed to an overloaded method is now its value type rather than a std::pair, and arrays, nested subscripts and operator[] work there too.

Validation

• New test __tests__/cpp-element-access-calls.test.ts (5 tests): all 5 fail on main and pass here.

• Mutations: switching off each part fails its test:

  • subscript typing: 6 tests, including the existing overload test for keys[i];
  • subscript receivers: 4;
  • operator[]: 1;
  • a map's mapped value: 3;
  • array dimensions: 2;
  • locals' dimensions: 2;
  • data members' dimensions: 1;
  • dimensions through an alias: 1;
  • a subscript of an array: 2;
  • *arr: 1;
  • a range-for over an array: 1;
  • overload arguments by subscript: 2.

• A/B (full index, main vs fix) on google/leveldb, fmtlib/fmt, redis (C control), facebook/rocksdb and protocolbuffers/protobuf.

  • An instrumented arm logged the decision behind every changed site. Every site was triaged by category, then line by line for each high-confidence removal and a sample of every target class.
  • rocksdb, 513 removed: library calls on elements that had been guessed onto project methods.
    • By method: size 191, push_back 60, c_str 45, reset 25, empty 21, …
    • The elements are std::string, std::vector, std::set, std::map, std::deque and std::unique_ptr.
    • For example, key.size() on const auto& key = keys_[ki] over a std::vector<std::string> reached Slice::size at 0.9.
    • Also 5 calls on the caller's own template parameter, like values[i].size() over a std::vector<TValue>.
  • rocksdb, 292 added and 843 retargeted:
    • Targets: Slice 256, WideColumn 124, PinnableAttributeGroup 102, IteratorBase 88, PinnableWideColumns 75, ColumnFamilyHandle 61, Status 53, DB 47, WalFile 31, TruncatedRangeDelIterator 26, autovector 25, …
    • autovector is reached for mems_list[k]->size() over an autovector of autovector pointers.
    • Some calls now take the InternalKey overload of InternalKeyComparator::Compare, for bottom_files[i]->largest arguments.
  • rocksdb, 807 metadata-only: same targets, confidence up to 0.9.
  • protobuf:
    • 58 removed. Of these, 23 are descriptor accessors like fields[i]->number() and extensions[0]->full_name(). FieldDescriptor declares them in its class and defines them with PROTOBUF_DEFINE_ACCESSOR macros, so no node exists for them. Every removed edge pointed at another class's namesake, like C#'s FieldGeneratorBase::number. The rest are library calls.
    • 5 added; 34 retargeted to the Java, C++ and Objective-C generator bases, ZeroCopyInputStream, DescriptorDatabase and EnumValueDescriptor; 66 metadata-only.
  • leveldb:
    • 12 removed: files_[level].size() over std::vector<FileMetaData*> files_[kNumLevels], which had been guessed as Slice::size.
    • 14 added: shard_[s].Insert(…) over LRUCache shard_[kNumShards], IteratorWrapper, files[i]->Read(…).
    • 1 retargeted, 11 metadata-only.
  • fmt:
    • 6 removed (library);
    • 4 added and 10 retargeted, to vendored gmock's MatcherBase, MatcherDescriberInterface and TestEventListener;
    • 2 metadata-only.
  • redis: byte-identical.

• Arrays: the first version read a class without an indexed operator[] as an array of itself. The A/B showed where that misfires:

  • std::vector<FileMetaData*> files_[kNumLevels] read one level too deep;
  • PinnableSlice values[2] read as chars through Slice::operator[].

  So arrays now come from the declarator instead. Between the two versions, 35 sites in leveldb and rocksdb changed, each toward the declared type.

• Full suite: 7,160 passed, 0 failed (82 skipped).

Remaining, out of scope:

• m.at(k).Foo() and v.front().Foo() chains, which need library containers' element-returning members;
• this->arr_[i].Foo(), which doesn't occur in the five repos.

🤖 Generated with Claude Code