fix(async): linearize await inside an async-generator finally (#8715)

An `await` inside a `finally` of an `async function*` compiled to a blocking
busy-wait instead of an async suspend — the finally analog of the #8681
await-in-catch deadlock fixed by #8707.

When a `try` in an async generator has a `finally` that yields or awaits, the
finally is linearized into its own dispatch states. `.next()` and `.throw()`
drive those states through the shared `__agstep` async-step driver, so a
finally `await` suspends on the microtask queue via `AsyncStepChain`. The
`.return()` closure, however, re-drove the SAME states through a separate
`build_dispatch_while_body(states, /*async_step*/ false, …)` continuation loop,
whose `StateExit::Await` lowering emits the busy-wait fallback
`__sent = await value; continue` (fs_await.rs → `js_wait_for_event`). So a
`.return()` that ran the finally — an early `break` in a `for await`, or an
explicit `gen.return(v)` while suspended in the try — block-waited on the
finally's `await`, monopolising the single runtime thread while the driver that
would settle it sits suspended, and deadlocked.

Fix: `.return()` no longer builds or runs an async_step=false loop for async
generators. After `build_abrupt_routing` records the pending return and jumps to
`finally_entry_state`, `.return()` hands off to the shared `__agstep` driver with
a fresh non-error resume (`AsyncGenResume(__agstep, undefined, false)`), exactly
as `.next()`/`.throw()` already do. `__agstep` dispatches from
`finally_entry_state`, runs the finally (its `yield`s settle this `.return()`'s
promise; its `await`s suspend on the microtask queue), and its completion-check
state re-raises the pending return as `{value, done: true}`. Sync generators are
unchanged — they have no `await` states, so their inline busy-wait clone stays
correct, and their `.return()` is a plain (non-driver) closure.

The `async_generator_linearizes_every_await_position` test re-adds the
`await-in-finally` case #8707 had removed (pointing here), plus
await-in-try-and-finally, await-in-try-catch-finally, and
yield-in-finally-with-await; all now leave zero residual `Expr::Await`.
Behaviorally verified byte-identical to Node v26 for explicit `.return()`,
`.throw()`, yield-in-finally, and try/catch/finally shapes, with no deadlock.

Full `cargo test -p perry-transform` is green.

Claude-Session: https://claude.ai/code/session_01TwxRkALrR9HKSF1zKLSTAF

Author: Ralph Küpper

changelog.d/8715-async-gen-finally-await.md

@@ -0,0 +1,15 @@
+fix(async): an `await` inside a `finally` of an `async function*` no longer
+compiles to a blocking busy-wait. When a `try` in an async generator has a
+finally that yields or awaits, the finally is linearized into its own dispatch
+states, and `.next()`/`.throw()` drive those states through the shared async-step
+driver so their `await`s suspend on the microtask queue. The `.return()` closure,
+however, re-drove the same states through a separate busy-wait dispatch loop
+(`__sent = await value; continue`) — so a `.return()` that ran the finally (an
+early `break` in a `for await`, or an explicit `.return()`) block-waited on the
+finally's `await`, monopolising the single runtime thread and deadlocking. This
+is the finally analog of the #8681 `await`-in-`catch` deadlock.
+
+`.return()` now hands the continuation off to the shared `__agstep` driver
+(a fresh non-error resume) after routing the pending return into the finally,
+exactly as `.next()`/`.throw()` already do, so a finally `await` suspends
+instead of blocking. Behavior for a finally that only yields is unchanged.

crates/perry-transform/src/async_to_generator_tests.rs

@@ -576,16 +576,49 @@ fn async_generator_linearizes_every_await_position() {
                 finally: None,
             }],
         ),
-        // NOTE: `await` inside a `finally` of a REAL async generator
-        // (`async function*`) is a SEPARATE, pre-existing gap in the
-        // `#4438` B2-finally lowering — the yielding finally's states are
-        // built with a raw `Expr::Await` instead of an async suspend, so it
-        // block-waits the same way. It is NOT addressed by this PR (which
-        // fixes the `was_plain_async` catch path); the closure test
-        // `async_closure_rewrite_leaves_no_residual_await` DOES cover
-        // `in-finally` for the `was_plain_async` path, which is clean.
-        // Tracked separately in #8715; omitted here so this test asserts
-        // only what this change fixes.
+        // #8715: `await` inside a `finally` of a REAL async generator
+        // (`async function*`). The yielding finally is linearized into its own
+        // dispatch states, but the `.return()` closure used to re-drive them
+        // through an async_step=false busy-wait loop (`__sent = await v;
+        // continue`) — a blocking wait, the finally analog of the #8681 catch
+        // deadlock. `.return()` now delegates the continuation to the shared
+        // `__agstep` driver, so the finally `await` suspends on the microtask
+        // queue and no raw `Expr::Await` survives.
+        (
+            "await-in-finally",
+            vec![Stmt::Try {
+                body: vec![y(Expr::Integer(0))],
+                catch: None,
+                finally: Some(vec![Stmt::Expr(await_(Expr::Integer(1)))]),
+            }],
+        ),
+        (
+            "await-in-try-and-finally",
+            vec![Stmt::Try {
+                body: vec![Stmt::Expr(await_(Expr::Integer(0))), y(Expr::Integer(5))],
+                catch: None,
+                finally: Some(vec![Stmt::Expr(await_(Expr::Integer(1)))]),
+            }],
+        ),
+        (
+            "await-in-try-catch-finally",
+            vec![Stmt::Try {
+                body: vec![y(Expr::Integer(0))],
+                catch: Some(CatchClause {
+                    param: None,
+                    body: vec![Stmt::Expr(await_(Expr::Integer(1)))],
+                }),
+                finally: Some(vec![Stmt::Expr(await_(Expr::Integer(2)))]),
+            }],
+        ),
+        (
+            "yield-in-finally-with-await",
+            vec![Stmt::Try {
+                body: vec![y(Expr::Integer(0))],
+                catch: None,
+                finally: Some(vec![y(Expr::Integer(8)), Stmt::Expr(await_(Expr::Integer(9)))]),
+            }],
+        ),
         (
             "await-in-if-inside-try-inside-loop",
             // The pi #6728 shape: await buried in nested control flow.

crates/perry-transform/src/generator/lower.rs

@@ -516,13 +516,19 @@ pub fn transform_generator_function_with_extra_captures(
     // #4374: clone the state-dispatch loop so the .throw() closure can
     // *continue* the state machine after running a catch handler.
     let while_body_for_throw = while_body.clone();
-    // #4438 B2-finally: the `.return()` closure needs the same continuation loop
-    // when it routes into a yielding finally (so the finally's `yield`s suspend).
-    // #6709: the `.return()` closure is NOT an async-step driver (it cannot
-    // chain an inner `await` through `CurrentStepClosure`), so its dispatch
-    // keeps the busy-wait `await` shape — matching pre-#6709 `.return()`.
+    // #4438 B2-finally: a `.return()` that routes into a yielding finally must
+    // keep driving the state machine so the finally's `yield`s/`await`s run.
+    // #8715: async generators delegate that continuation to the shared `__agstep`
+    // step driver (see the `has_yielding_finally` branch below), so an `await`
+    // inside the finally suspends on the microtask queue via `AsyncStepChain`
+    // exactly as it does on the `.next()`/`.throw()` paths. Building an
+    // async_step=false dispatch loop here instead would lower every such `await`
+    // to a blocking busy-wait (`__sent = await v; continue`) — the finally analog
+    // of the #8681 catch deadlock — so async generators build none. Sync
+    // generators keep the busy-wait clone: they have no `await` states, so it
+    // stays correct, and their `.return()` is a plain (non-driver) closure.
     let while_body_for_return = if is_async_generator {
-        build_dispatch_while_body(&states, false, state_id, done_id, sent_id)
+        Vec::new()
     } else {
         while_body.clone()
     };
@@ -577,7 +583,10 @@ pub fn transform_generator_function_with_extra_captures(
     } else {
         while_body_for_throw
     };
-    let while_body_for_return = if wrap_dispatch {
+    // #8715: async generators no longer run a local `.return()` dispatch loop
+    // (`while_body_for_return` is empty — they delegate to `__agstep`), so skip
+    // wrapping it. Sync generators still wrap their busy-wait clone.
+    let while_body_for_return = if wrap_dispatch && !is_async_generator {
         let disp_err_id = alloc_local(next_local_id);
         wrap_dispatch_loop(
             while_body_for_return,
@@ -977,10 +986,10 @@ pub fn transform_generator_function_with_extra_captures(
         ))));
         if has_yielding_finally {
             // #4438 B2-finally: route `.return(v)` into the innermost enclosing
-            // yielding finally (record the pending return + jump in), then fall
-            // through to the continuation loop so the finally's `yield`s suspend;
-            // its completion check re-raises the return. Catches don't catch a
-            // return completion, so only finally routes apply.
+            // yielding finally — record the pending return and jump to
+            // `finally_entry_state`. Catches don't catch a return completion, so
+            // only finally routes apply; on no match, `return_fallback` completes
+            // the generator directly (never reaching the continuation below).
             return_resume_body.extend(build_abrupt_routing(
                 &catches,
                 &finallys,
@@ -994,10 +1003,36 @@ pub fn transform_generator_function_with_extra_captures(
                 false,
                 return_fallback,
             ));
-            return_resume_body.push(Stmt::While {
-                condition: Expr::Bool(true),
-                body: while_body_for_return,
-            });
+            if is_async_generator {
+                // #8715: a matched route has set `state = finally_entry_state`
+                // and recorded the pending return in the shared boxed locals.
+                // Hand off to the shared `__agstep` driver (a fresh, non-error
+                // resume) rather than run a local async_step=false loop, so a
+                // finally `await` suspends on the microtask queue (`AsyncStepChain`
+                // re-entering `__agstep`) instead of block-waiting — the fix for
+                // this issue. `__agstep` dispatches from `finally_entry_state`,
+                // runs the finally (its `yield`s settle this `.return()`'s
+                // promise, its `await`s suspend), and its completion-check state
+                // re-raises the pending return as `{value, done: true}`. This
+                // mirrors how `.next()`/`.throw()` already drive a yielding
+                // finally. `wrap_generator_resume_body` clears `executing` before
+                // this return, so `__agstep`'s re-entrancy guard passes.
+                let agstep_local_id =
+                    agstep_id.expect("agstep_id is set for async generators");
+                return_resume_body.push(Stmt::Return(Some(Expr::AsyncGenResume {
+                    step_closure: Box::new(Expr::LocalGet(agstep_local_id)),
+                    value: Box::new(Expr::Undefined),
+                    is_error: false,
+                })));
+            } else {
+                // Sync generators re-drive the finally inline in this closure —
+                // no microtask suspend is needed (they have no `await`), and the
+                // finally's `yield`s return `{value, done: false}` directly.
+                return_resume_body.push(Stmt::While {
+                    condition: Expr::Bool(true),
+                    body: while_body_for_return,
+                });
+            }
         } else {
             return_resume_body.extend(return_fallback);
         }