[colbymchenry]

Follow-up to #2701: Rust calls chained on .as_mut() / .as_ref().

What was wrong

A call written x.as_mut().m(…) reaches the resolver by its bare name (m), so it was matched by name alone — #2701's Pin lookup never saw it. On the A/B repos that linked, among others:

call	receiver declared	main linked it to
tokio self.delay.as_mut().reset(now) ×3	delay: Pin<Box<Sleep>>	the calling Interval::reset
tokio this.timer.as_mut().as_pin_mut() ×4	Option<…>	LinesStream::as_pin_mut
tokio reader.as_mut().poll_read(…), f.as_mut().poll(cx) ×6	Pin<&mut R>, a generic F	BufReader::poll_read, BlockingTask::poll
tokio entry.as_ref().might_be_registered() ×2	NonNull<TimerShared>	StateCell::might_be_registered
dioxus self.child.as_mut().map(…), self.server.as_mut().filter(…), x.as_ref().context(…) ×12	Option<…>	GenerationalRefMut::map, Store::filter, DiffState::context, …

The change (src/resolution/name-matcher.ts)

matchRustAsMutChainCall, an exclusive hook beside the C++ chain hooks in matchReferenceInner, reads the chain from the call's own column (self.delay.as_mut().reset(, with an optional ? after as_mut()) and resolves the method on what as_mut / as_ref hands back for the receiver's declared type (rustLentCall):

• a Pin<P> re-pins what it points at → fix(rust): a call through a Pin, Rc or Box self reaches the method Rust picks, inherent first #2701's lookup through Pin (rustPinnedCallOn);
• a Box, NonNull, Rc, Arc or Cow lends what it holds → that type's method (rustCallOnType);
• an Option / Result stays one — std's methods, no edge — unless ? takes the value back out (self.client.as_mut()?.next());
• a type parameter, or another std or dependency type → no project method;
• a project type's own as_mut (return type not read here), or a receiver whose type isn't written where it can be read (a pin! local, a call's result) → undefined, the name strategies decide as before.

The receiver's type comes from: self (through self: Pin<&mut Self>, self.as_mut() is that same Pin); a field's declaration in its struct (rustFieldDeclaration), also when a pin_project! macro wraps the struct and through pin-project's let me = self.project(), where a #[pin] field is Pin<&mut T>; a local's or parameter's declaration (rustDeclaredTypeText, which now also reads let x: T = … — #2701's reader stopped at the let and missed the annotation). A cheap line check (.as_mut() / .as_ref() on the call's line) keeps the hook off every other bare call.

Validation

Fresh full indexes, main f2f5dee vs this branch, keyed by call site (re-run after rebasing onto 90e2456, which brought #2707's Rust extractor change: the same changed sites, site for site; the branch now sits on a31bf2f, whose one newer commit is C-only):

repo	removed	retargeted	added	same target, now by declared type
tokio	15	6	13	27
dioxus	16	1	0	1
ripgrep	1	0	0	0
gleam	1	0	0	0
axum, clap, serde	—	—	—	identical
express, gin, flask, leveldb (controls)	—	—	—	byte-identical dumps

Every changed site was read:

• removed (33), all wrong on main: Option / Result methods (map, filter, cloned, is_none_or, as_pin_mut, anyhow's context) guessed onto project types; generic readers, writers, sinks and futures (R, W, S, F); Pin::get_ref → AsyncFd::get_ref; an external Stream / dyn Future / web_sys::Node method; ripgrep's path.as_ref().metadata() and gleam's path.as_ref().replace(…) on P: AsRef<Path> / a string.
• retargeted (7), all now right: Sleep::reset ×3 (was the calling Interval::reset), TimerShared::might_be_registered ×2 (was StateCell's), Fuse::poll_next (was the calling Peekable::poll_next), dioxus HasAnimationData::as_any through a Box<dyn HasAnimationData> (was SerializedAnimationData::as_any).
• added (13), all right: pin-projected fields (MaybeDone::poll ×3, Root::poll ×3, Notified::poll, MaybeDangling::poll, Fuse::poll_next), NonNull<TimerShared> calls ×3, Sleep::reset.

Not in this PR: an earlier draft also made the shared Rust/Go call-shape reader stop treating a receiver that merely starts with the method's name (builder().build(), abort_handle…abort()) as a bare call. That fixed a few calls but exposed others to the name guess (dioxus's builder().build() → BuildRequest::build), so it was dropped; the as_mut hook reads its chain itself instead.

Cost: none measurable. The hook's own time, from the resolver's stage profiler (CODEGRAPH_RESOLVE_PROFILE=2) summed over every worker: 180 ms on tokio (8,014 bare calls checked), 61 ms on dioxus (12,609), against 31–36 s of CPU per full index; a call without .as_mut() / .as_ref() on its line costs one regex test. Interleaved full indexes (CPU, median of 3) moved within this machine's 2–4 s run-to-run spread: tokio 31.2 s → 32.1 s, dioxus 34.2 s → 36.1 s.

Tests

__tests__/rust-as-mut-chains.test.ts (7 cases over Pin<Box<T>>, Option<T>, Box<T> and NonNull<T> fields, Option through ?, pin-project fields pinned and not, Pin<&mut R> generic and named, self.as_mut() through self: Pin<&mut Self> with Pin::get_mut): 5 fail on main, the other 2 guard behavior main already has. With #2701's and the other Rust suites: 39/39. Full suite (CODEGRAPH_KERNEL_EXPECT=1, kernel built, 4 workers): 7619 passed, 0 failed, 80 skipped.

Overlap

The peer session working on Rust receivers typed by their declaration (branch claude/rust-receiver-types) handles ident.method refs only and confirmed it doesn't see these bare chained calls; the two touch different functions.

🤖 Generated with Claude Code