[colbymchenry]

Summary

In Go, a method called on a variable that holds what a call returns, like r.Reset() after r := newRing(), v1.GET(…) after v1 := router.Group("/v1"), or r.Use(…) after c, r := CreateTestContext(w), now links to the method of the type that call returns. Before, these were matched by the method's name alone. A value of a type from outside the project, such as c, _ := net.Dial(…), an *os.File or an http.Handler a project function returns, now links to nothing instead of a project namesake.

This reimplements @danusha2345's PR #1954 on the value typer main already has (#2782), with their tests.

Root cause

inferLocalReceiverType types a Go receiver from the type its in-scope declaration writes (s *store.Store, x := &T{…}) and from a few line patterns. None of them reads a call's result, so a local bound from a call fell through to name matching. That guess took a namesake: etcd's lg := s.Logger() linked lg.Info(…) to a project logger's Info, and prometheus's NewWriteHandler(…) results to the read handler's ServeHTTP. Or it found nothing.

Fix

main has typed a value from its declaration since #2782 (goDeclaredValueType → goCalledType / goSignatureResultType / goMethodResultType), but only for indexed and chained calls. PR #1954 added a second, regex-based derivation of the same thing (goBoundCallResult with a paren scan of up to 40 lines per scanned line, goFirstResult, goMethodOfWrittenType). This change uses main's typer for plain receivers instead:

• A local whose in-scope declaration has a call or a conversion as its value takes that value's type: the k-th result of several, a method's result, or the conversion's type. The type is read in the file that writes it (GoReceiverDecl.file), for matchMethodCall, matchMemberFunctionRef and matchGoFieldChainCall.
• A value of a type from outside the project links nothing. That includes a project function's result written as an outside interface: about 80% of those by-name links were wrong in the A/B below.
• Assertions stay with goAssertedLocalType, and a type switch's guard is not typed from its value.

Changes to the typer, for every path that uses it:

• a trailing .(T) is a T whatever its prefix (prometheus's rw.converterPool.Get().(*PrometheusConverter));
• a call through a func-typed field gives the field's result type (harbor's h.ArtifactControllerFunc());
• a bare call that a dot import brings in is typed (kubernetes' package watch_test calling NewStreamWatcher(…));
• a method of a struct or interface literal, of an outside generic type (grpc's ServerStreamingClient[Resp].Recv()), or one the index doesn't show on a project type gives an unknown type rather than "outside", so name matching runs as before.

Validation

• Red→green: go-receiver-type.test.ts (the PR's 23 cases plus 4 of ours), go-std-method-receivers.test.ts and go-chained-receivers.test.ts: on main 19 of 51 fail; all pass here. All 45 go-*.test.ts files pass (491 tests).

• Expectations changed, where a call is now typed:

  • the PR's selfArg: the inner e of if e := wrap(e.Run()) is wrap's *ringLog;
  • the PR's chainedCtor: r := newRing().Reader() is a reader;
  • versionText and urls (std-method file): typed at 0.9;
  • Go: factory calls and GORM-style chains resolve to unrelated symbols #2543's h.get().Run() through a get func() A field: A::Run;
  • the PR's interfaceResult now links nothing (policy above).

• A/B (same kernel; main 2577601 for this branch; the PR on its own cherry-pick; no Go drift between the bases):

  repo	PR vs main (added / removed / moved)	this branch vs main	this branch vs PR
  gin	+153 / -48 / 65	+194 / -52 / 65	+41 / -4 / 0
  etcd	+3,468 / -225 / 395	+4,441 / -359 / 743	+976 / -137 / 348
  prometheus	+3,442 / -166 / 1,027	+7,041 / -239 / 1,409	+3,658 / -132 / 400
  kubernetes staging subtree (1,247 Go files)	+992 / -68 / 341	+1,154 / -86 / 354	+194 / -50 / 13
  harbor without its portal (1,611 Go files)	+333 / -104 / 355	+423 / -138 / 366	+93 / -37 / 11

  express and flask are byte-identical.

  Triage, by sample:

  • Links this branch adds over the PR are right (e.g. clus := testRunner.NewCluster(…), app := head.Appender(ctx), tx := b.BatchTx(), v1 := router.Group(…)).
  • etcd's retargets are this branch right, 14 of 14 (typed vs the PR's by-name).
  • The PR's edges this branch drops are mostly wrong guesses (gin 5/5, etcd 13/14, prometheus 10/12).
  • Its own removals are outside values (zap, os, grpc, otel, beego, pflag, structured-merge-diff).
  • The PR's own removals lost right links on range variables, field values and two-dot callees (prometheus 59, k8s 31); this branch leaves those to main's scan.

• Outside-interface results (a project function's result written as an outside type), about 167 by-name links over the five repos, about 80% wrong:

  • wrong: grpc streams linked to a proxy adapter's Recv; every NewWriteHandler(…) result linked to readHandler::ServeHTTP; a reflect.Value linked to QuantityValue::Type; harbor's HTTPStatusCodeHealthChecker(…), docker's health.CheckFunc, linked to updater::Check;
  • right: factories that return their own type, like harbor's PeriodicHealthChecker (updater), kubernetes' NewTracker (TrackerValue), etcd's NewRaftLoggerFromZapCore (zapRaftLogger).
  • Follow-up, not in this PR: read the factory's return &T{…} to type that minority.

• Known loss, as in fix(go): a call through a variable declared with another package's type reaches that type's method #2706: a type defined over an interface (type Codec Serializer) has no methods in the index, so a coincidental guess on such a value (codec := ….LegacyCodec(gv)) goes.

• The ~199 kubernetes call-result sites fix(go): a call on an indexed element reaches its element type's method #2755 left are indexed calls, which fix(go): a call further along a chain links the method of the type the chain reaches #2782 already serves: no change (etcd 24, prometheus 161, k8s 48, harbor 24 sites; 0 differ).

• CPU, profile self time of receiver typing, main → this branch: prometheus 1.82 → 1.72 s, kubernetes subtree 0.88 → 0.92 s, harbor 1.02 → 0.86 s. For comparison, the PR's goBoundCallResult alone took 390–699 ms on these.

• Full suite on the exact commit (Windows, kernel built, CODEGRAPH_KERNEL_EXPECT=1, --maxWorkers=4): 777 files, 9,202 passed, 0 failed, 85 skipped.

Reimplements #1954 by @danusha2345, whose tests are included (co-author on the commit). Its overlap with first-time PRs: #1640's cases already pass on main (#2782), and #1521 (multi-module layouts) is outside this change.

🤖 Generated with Claude Code