fix(cpp): a call of a member template with explicit template arguments has the type the template returns

[colbymchenry]

Summary

An auto local initialized from a call of a C++ member template with explicit template arguments got no type:

• protocolbuffers/protobuf: const auto& rep = reflection->GetRaw<RepeatedPtrField<CPPTYPE>>(message, field); and auto* lhs = r->MutableRaw<InlinedStringField>(…);
• facebook/rocksdb: auto bbto = factory->GetOptions<BlockBasedTableOptions>() and auto del_factory = factory->CheckedCast<CompactOnDeletionCollectorFactory>().

There were two reasons:

• cppExpressionType read a member call only when its callee ended in the method's name, so the explicit <…> made it no call at all.
• These templates are declared to return their own template parameter: template <typename Type> const Type& GetRaw(…), T* GetOptions(), T* CheckedCast(). Read as written, that return type is template-dependent.

Calls on such locals fell to name matching: rep.size() guessed SooRep::size, and lhs_string->InternalSwap(…) reached MicroString or ArenaStringPtr by luck.

This was found while working on #2555. protobuf's case is the second shape that chip named, reflection_visit_fields.h:142.

What changed

All in name-matcher.ts:

• cppExpressionType: a member call may write explicit template arguments (recv->method<Args>(…)). Such a call only reaches template overloads; a non-template Widget* Find() beside template <typename T> T* Find() is never what it calls.
• cppMemberTemplateReturn: when every template overload is declared to return the same one of its own parameters, perhaps const, a pointer or a reference, the call returns the argument for it, so qualified. It is read as written at the call. Otherwise the return type is read as written, as before. That still serves a project class template like RepeatedFieldRef<T> or Ptr<T>.
• cppOwnTemplateParameters: reads a function template's own template <…> header, in order and without defaults. A template's node starts at its declaration in both engines, so the header is the code right before it.
• cppStatementFrom: a macro body's \-spliced lines are read as one statement.

Validation

Site-keyed edge diffs of main e03ff673 against this branch. The base arm is the integration build #2554 and #2555 were merged from, whose tree equals e03ff673; both arms were built from one tree, kernel included.

repo	removed	added	retargeted	metadata-only
google/leveldb	0	0	0	0
fmtlib/fmt	0	0	0	0
protocolbuffers/protobuf	3	5	12	1
facebook/rocksdb	1	6	0	22
redis/redis (C, control)	0	0	0	0

Node and file counts are identical. Every change was triaged against the initializer in the source, and all 50 are right.

protobuf:

• SwapFieldHelper (generated_message_reflection.cc), 17 sites. Each receiver is a MutableRaw<T>(…) local, and its calls now reach T's methods:
  • lhs_cord → RepeatedField<absl::Cord>: InternalSwap, Swap (added);
  • lhs_string → RepeatedPtrFieldBase, InlinedStringField and MicroString in turn (retargeted from MicroString/RepeatedPtrFieldStringAccessor guesses at 0.65). In SwapStringField it had reached ArenaStringPtr at 0.9, through an earlier function's ArenaStringPtr* lhs_string;
  • lhs_map → MapFieldBase (retargeted from UntypedMapBase; Swap from 0.65 to 0.9);
  • lhs_rm → RepeatedPtrFieldBase (added).
• reflection_visit_fields.h, 2 sites. rep.size() reaches RepeatedPtrField::size, and map.size() reaches MapFieldBase::size. They had reached SooRep and java::FieldGeneratorMap. This is the macro case, read with its \ splices.
• arena_unittest.cc. outside.get() on auto outside = arena.Make<int>(), a Ptr<T>, reaches Arena::Ptr::get.
• Removed (3). rep_str.size() on GetRepeatedFieldRef<std::string>(…), a RepeatedFieldRef<T> whose size its specializations declare, no longer guesses SooRep::size.

rocksdb:

• Metadata-only (22). CheckedCast<T>() locals (del_factory, bfp, collector_factory, merge_operator, adapter) keep their targets, now found at 0.9 instead of 0.65–0.7.
• Added (6).
  • comp->Inner() on env->CheckedCast<CompositeEnvWrapper>();
  • config->GetOption(…) on opt_info->AsRawPointer<Configurable>(…);
  • the two ChanglingMergeOperator / ChanglingCompactionFilterFactory Name() calls on two-line CheckedCast initializers;
  • ttl_mo->CheckedCast<…>() on a CheckedCast<TtlMergeOperator>() local.
• Removed (1). shared->get() on GetOptions<std::shared_ptr<Configurable>>(…), a pointer to a std::shared_ptr, no longer guesses SharedCleanablePtr::get.

Tests

__tests__/cpp-member-template-returns.test.ts, 3 tests:

• protobuf's const auto& rep = over a \-spliced reflection->GetRaw<RepeatedPtrField<int>>(…) reaches RepeatedPtrField::size, with an out-of-line GetRaw definition and a namesake SooRep::size present. The extractor doesn't parse #define bodies, so a fixture can't use protobuf's macro itself; the A/B above covers that site.
• rocksdb's auto bbto = factory->GetOptions<BlockBasedTableOptions>() against its const and non-const overloads.
• Overload guards:
  • Find<Gadget>() beside a non-template Widget* Find();
  • Make<Gadget>() with a defaulted template <typename T = Widget>;
  • Get<Widget, Gadget>(1) against overloads returning different parameters, left unguessed.

Mutation check (each guard broken alone; every one fails a test):

guard	fails
a member call may write explicit template arguments	all three
explicit template arguments call only a template	Find<Gadget>()
a parameter return type takes the argument	all three
every overload returns the same parameter	Get<Widget, Gadget>(1)
a defaulted parameter is read by its name	Make<Gadget>()
the header right before the node	all three
a macro body's lines are spliced at their \	the GetRaw local

The template keyword check in the header reader recognizes syntax; no output depends on it alone.

One existing test changed: cpp-auto-local-receiver.test.ts, "an initializer that goes on past its line is read to its end". It expected ChanglingMergeOperator::SetName at 0.7, reached by a fallback because CheckedCast<T>()'s T* couldn't be read. It now gets the same target at 0.9, from the type CheckedCast<test::ChanglingMergeOperator>() returns, as the A/B did at options_test.cc:3960.

Full suite: 7,062 passed, 82 skipped. The one failure in the first run was that changed test, at its old 0.7, and it passes with its new expectation.

🤖 Generated with Claude Code