wasmtime/runtime/component/

concurrent.rs

//! Runtime support for the Component Model Async ABI.
//!
//! This module and its submodules provide host runtime support for Component
//! Model Async features such as async-lifted exports, async-lowered imports,
//! streams, futures, and related intrinsics.  See [the Async
//! Explainer](https://github.com/WebAssembly/component-model/blob/main/design/mvp/Concurrency.md)
//! for a high-level overview.
//!
//! At the core of this support is an event loop which schedules and switches
//! between guest tasks and any host tasks they create.  Each
//! `Store` will have at most one event loop running at any given
//! time, and that loop may be suspended and resumed by the host embedder using
//! e.g. `StoreContextMut::run_concurrent`.  The `StoreContextMut::poll_until`
//! function contains the loop itself, while the
//! `StoreOpaque::concurrent_state` field holds its state.
//!
//! # Public API Overview
//!
//! ## Top-level API (e.g. kicking off host->guest calls and driving the event loop)
//!
//! - `[Typed]Func::call_concurrent`: Start a host->guest call to an
//! async-lifted or sync-lifted import, creating a guest task.
//!
//! - `StoreContextMut::run_concurrent`: Run the event loop for the specified
//! instance, allowing any and all tasks belonging to that instance to make
//! progress.
//!
//! - `StoreContextMut::spawn`: Run a background task as part of the event loop
//! for the specified instance.
//!
//! - `{Future,Stream}Reader::new`: Create a new Component Model `future` or
//! `stream` which may be passed to the guest.  This takes a
//! `{Future,Stream}Producer` implementation which will be polled for items when
//! the consumer requests them.
//!
//! - `{Future,Stream}Reader::pipe`: Consume a `future` or `stream` by
//! connecting it to a `{Future,Stream}Consumer` which will consume any items
//! produced by the write end.
//!
//! ## Host Task API (e.g. implementing concurrent host functions and background tasks)
//!
//! - `LinkerInstance::func_wrap_concurrent`: Register a concurrent host
//! function with the linker.  That function will take an `Accessor` as its
//! first parameter, which provides access to the store between (but not across)
//! await points.
//!
//! - `Accessor::with`: Access the store and its associated data.
//!
//! - `Accessor::spawn`: Run a background task as part of the event loop for the
//! store.  This is equivalent to `StoreContextMut::spawn` but more convenient to use
//! in host functions.

use self::error_contexts::GlobalErrorContextRefCount;
use crate::component::func::{Func, call_post_return};
use crate::component::{
    HasData, HasSelf, Instance, Resource, ResourceTable, ResourceTableError, RuntimeInstance,
};
use crate::fiber::{self, StoreFiber, StoreFiberYield};
use crate::hash_set::HashSet;
#[cfg(feature = "gc")]
use crate::module::ModuleRegistry;
use crate::prelude::*;
use crate::store::{Store, StoreId, StoreInner, StoreOpaque, StoreToken};
#[cfg(feature = "gc")]
use crate::vm::GcRootsList;
use crate::vm::component::{CallContext, ComponentInstance, CurrentScope, InstanceState, Scope};
use crate::vm::{
    AlwaysMut, SendSyncPtr, UncaughtException, VMFuncRef, VMLazyThread, VMMemoryDefinition, VMStore,
};
use crate::{AsContext, AsContextMut, Result, StoreContext, StoreContextMut, ValRaw, bail};
use crate::{Instance as ModuleInstance, bail_bug};
use alloc::borrow::ToOwned;
use alloc::collections::btree_map::Entry;
use alloc::collections::{BTreeMap, BTreeSet, VecDeque};
use core::any::Any;
use core::cell::UnsafeCell;
use core::fmt;
use core::future;
use core::future::Future;
use core::marker::PhantomData;
use core::mem::{self, ManuallyDrop, MaybeUninit};
use core::ops::DerefMut;
use core::pin::{Pin, pin};
use core::ptr::{self, NonNull};
use core::task::{Context, Poll, Waker};
use futures::channel::oneshot;
use futures::stream::{FuturesUnordered, StreamExt};
use futures_and_streams::{FlatAbi, ReturnCode, TransmitHandle, TransmitIndex};
use table::{TableDebug, TableId};
use wasmtime_environ::component::{
    CanonicalAbiInfo, CanonicalOptions, CanonicalOptionsDataModel, MAX_FLAT_PARAMS,
    MAX_FLAT_RESULTS, OptionsIndex, PREPARE_ASYNC_NO_RESULT, PREPARE_ASYNC_WITH_RESULT,
    RuntimeComponentInstanceIndex, RuntimeTableIndex, StringEncoding,
    TypeComponentGlobalErrorContextTableIndex, TypeComponentLocalErrorContextTableIndex,
    TypeFutureTableIndex, TypeStreamTableIndex, TypeTupleIndex,
};
use wasmtime_environ::packed_option::ReservedValue;
use wasmtime_environ::{ModuleInternedTypeIndex, NUM_COMPONENT_CONTEXT_SLOTS, Trap};
#[cfg(feature = "gc")]
use wasmtime_unwinder::Unwind;

pub use abort::JoinHandle;
pub use func::{FuncCallConcurrent, TypedFuncCallConcurrent};
pub use future_stream_any::{FutureAny, StreamAny};
pub use futures_and_streams::{
    Destination, DirectDestination, DirectSource, ErrorContext, FutureConsumer, FutureProducer,
    FutureReader, GuardedFutureReader, GuardedStreamReader, ReadBuffer, Source, StreamConsumer,
    StreamProducer, StreamReader, StreamResult, VecBuffer, WriteBuffer,
};
pub(crate) use futures_and_streams::{ResourcePair, lower_error_context_to_index};
#[cfg(feature = "task-group-hook")]
pub use task_group_hook::TaskGroupHook;
pub use task_group_hook::TaskGroupId;

mod abort;
mod error_contexts;
mod func;
mod future_stream_any;
mod futures_and_streams;
pub(crate) mod table;
#[cfg(feature = "task-group-hook")]
mod task_group_hook;
#[cfg(not(feature = "task-group-hook"))]
mod task_group_hook_disabled;
#[cfg(not(feature = "task-group-hook"))]
use task_group_hook_disabled as task_group_hook;
pub(crate) mod tls;

/// Constant defined in the Component Model spec to indicate that the async
/// intrinsic (e.g. `future.write`) has not yet completed.
const BLOCKED: u32 = 0xffff_ffff;

/// Corresponds to `CallState` in the upstream spec.
#[derive(Clone, Copy, Eq, PartialEq, Debug)]
pub enum Status {
    Starting = 0,
    Started = 1,
    Returned = 2,
    StartCancelled = 3,
    ReturnCancelled = 4,
}

impl Status {
    /// Packs this status and the optional `waitable` provided into a 32-bit
    /// result that the canonical ABI requires.
    ///
    /// The low 4 bits are reserved for the status while the upper 28 bits are
    /// the waitable, if present.
    pub fn pack(self, waitable: Option<u32>) -> u32 {
        assert!(matches!(self, Status::Returned) == waitable.is_none());
        let waitable = waitable.unwrap_or(0);
        assert!(waitable < (1 << 28));
        (waitable << 4) | (self as u32)
    }
}

/// Corresponds to `EventCode` in the Component Model spec, plus related payload
/// data.
#[derive(Clone, Copy, Debug)]
enum Event {
    None,
    Subtask {
        status: Status,
    },
    StreamRead {
        code: ReturnCode,
        pending: Option<(TypeStreamTableIndex, u32)>,
    },
    StreamWrite {
        code: ReturnCode,
        pending: Option<(TypeStreamTableIndex, u32)>,
    },
    FutureRead {
        code: ReturnCode,
        pending: Option<(TypeFutureTableIndex, u32)>,
    },
    FutureWrite {
        code: ReturnCode,
        pending: Option<(TypeFutureTableIndex, u32)>,
    },
    Cancelled,
}

impl Event {
    /// Lower this event to core Wasm integers for delivery to the guest.
    ///
    /// Note that the waitable handle, if any, is assumed to be lowered
    /// separately.
    fn parts(self) -> (u32, u32) {
        const EVENT_NONE: u32 = 0;
        const EVENT_SUBTASK: u32 = 1;
        const EVENT_STREAM_READ: u32 = 2;
        const EVENT_STREAM_WRITE: u32 = 3;
        const EVENT_FUTURE_READ: u32 = 4;
        const EVENT_FUTURE_WRITE: u32 = 5;
        const EVENT_CANCELLED: u32 = 6;
        match self {
            Event::None => (EVENT_NONE, 0),
            Event::Cancelled => (EVENT_CANCELLED, 0),
            Event::Subtask { status } => (EVENT_SUBTASK, status as u32),
            Event::StreamRead { code, .. } => (EVENT_STREAM_READ, code.encode()),
            Event::StreamWrite { code, .. } => (EVENT_STREAM_WRITE, code.encode()),
            Event::FutureRead { code, .. } => (EVENT_FUTURE_READ, code.encode()),
            Event::FutureWrite { code, .. } => (EVENT_FUTURE_WRITE, code.encode()),
        }
    }
}

/// Corresponds to `CallbackCode` in the spec.
mod callback_code {
    pub const EXIT: u32 = 0;
    pub const YIELD: u32 = 1;
    pub const WAIT: u32 = 2;
}

// Flags passed to the `start-call` intrinsic.
const START_FLAG_ASYNC_CALLEE: u32 = wasmtime_environ::component::START_FLAG_ASYNC_CALLEE as u32;
const START_FLAG_ASYNC_CALLER: u32 = wasmtime_environ::component::START_FLAG_ASYNC_CALLER as u32;

/// Provides access to either store data (via the `get` method) or the store
/// itself (via [`AsContext`]/[`AsContextMut`]), as well as the component
/// instance to which the current host task belongs.
///
/// See [`Accessor::with`] for details.
pub struct Access<'a, T: 'static, D: HasData + ?Sized = HasSelf<T>> {
    store: StoreContextMut<'a, T>,
    get_data: fn(&mut T) -> D::Data<'_>,
}

impl<'a, T, D> Access<'a, T, D>
where
    D: HasData + ?Sized,
    T: 'static,
{
    /// Creates a new [`Access`] from its component parts.
    pub fn new(store: StoreContextMut<'a, T>, get_data: fn(&mut T) -> D::Data<'_>) -> Self {
        Self { store, get_data }
    }

    /// Get mutable access to the store data.
    pub fn data_mut(&mut self) -> &mut T {
        self.store.data_mut()
    }

    /// Get mutable access to the store data.
    pub fn get(&mut self) -> D::Data<'_> {
        (self.get_data)(self.data_mut())
    }

    /// Spawn a background task.
    ///
    /// See [`Accessor::spawn`] for details.
    pub fn spawn(&mut self, task: impl for<'fut> AccessorTask<'fut, T, D>) -> Result<JoinHandle>
    where
        T: 'static,
    {
        let accessor = Accessor {
            get_data: self.get_data,
            token: StoreToken::new(self.store.as_context_mut()),
        };
        self.store
            .as_context_mut()
            .spawn_with_accessor(accessor, task)
    }

    /// Returns the getter this accessor is using to project from `T` into
    /// `D::Data`.
    pub fn getter(&self) -> fn(&mut T) -> D::Data<'_> {
        self.get_data
    }
}

impl<'a, T, D> AsContext for Access<'a, T, D>
where
    D: HasData + ?Sized,
    T: 'static,
{
    type Data = T;

    fn as_context(&self) -> StoreContext<'_, T> {
        self.store.as_context()
    }
}

impl<'a, T, D> AsContextMut for Access<'a, T, D>
where
    D: HasData + ?Sized,
    T: 'static,
{
    fn as_context_mut(&mut self) -> StoreContextMut<'_, T> {
        self.store.as_context_mut()
    }
}

/// Provides scoped mutable access to store data in the context of a concurrent
/// host task future.
///
/// This allows multiple host task futures to execute concurrently and access
/// the store between (but not across) `await` points.
///
/// # Rationale
///
/// This structure is sort of like `&mut T` plus a projection from `&mut T` to
/// `D::Data<'_>`. The problem this is solving, however, is that it does not
/// literally store these values. The basic problem is that when a concurrent
/// host future is being polled it has access to `&mut T` (and the whole
/// `Store`) but when it's not being polled it does not have access to these
/// values. This reflects how the store is only ever polling one future at a
/// time so the store is effectively being passed between futures.
///
/// Rust's `Future` trait, however, has no means of passing a `Store`
/// temporarily between futures. The [`Context`](core::task::Context) type does
/// not have the ability to attach arbitrary information to it at this time.
/// This type, [`Accessor`], is used to bridge this expressivity gap.
///
/// The [`Accessor`] type here represents the ability to acquire, temporarily in
/// a synchronous manner, the current store. The [`Accessor::with`] function
/// yields an [`Access`] which can be used to access [`StoreContextMut`], `&mut
/// T`, or `D::Data<'_>`. Note though that [`Accessor::with`] intentionally does
/// not take an `async` closure as its argument, instead it's a synchronous
/// closure which must complete during on run of `Future::poll`. This reflects
/// how the store is temporarily made available while a host future is being
/// polled.
///
/// # Implementation
///
/// This type does not actually store `&mut T` nor `StoreContextMut<T>`, and
/// this type additionally doesn't even have a lifetime parameter. This is
/// instead a representation of proof of the ability to acquire these while a
/// future is being polled. Wasmtime will, when it polls a host future,
/// configure ambient state such that the `Accessor` that a future closes over
/// will work and be able to access the store.
///
/// This has a number of implications for users such as:
///
/// * It's intentional that `Accessor` cannot be cloned, it needs to stay within
///   the lifetime of a single future.
/// * A future is expected to, however, close over an `Accessor` and keep it
///   alive probably for the duration of the entire future.
/// * Different host futures will be given different `Accessor`s, and that's
///   intentional.
/// * The `Accessor` type is `Send` and `Sync` irrespective of `T` which
///   alleviates some otherwise required bounds to be written down.
///
/// # Using `Accessor` in `Drop`
///
/// The methods on `Accessor` are only expected to work in the context of
/// `Future::poll` and are not guaranteed to work in `Drop`. This is because a
/// host future can be dropped at any time throughout the system and Wasmtime
/// store context is not necessarily available at that time. It's recommended to
/// not use `Accessor` methods in anything connected to a `Drop` implementation
/// as they will panic and have unintended results. If you run into this though
/// feel free to file an issue on the Wasmtime repository.
pub struct Accessor<T: 'static, D = HasSelf<T>>
where
    D: HasData + ?Sized,
{
    token: StoreToken<T>,
    get_data: fn(&mut T) -> D::Data<'_>,
}

/// A helper trait to take any type of accessor-with-data in functions.
///
/// This trait is similar to [`AsContextMut`] except that it's used when
/// working with an [`Accessor`] instead of a [`StoreContextMut`]. The
/// [`Accessor`] is the main type used in concurrent settings and is passed to
/// functions such as [`Func::call_concurrent`].
///
/// This trait is implemented for [`Accessor`] and `&T` where `T` implements
/// this trait. This effectively means that regardless of the `D` in
/// `Accessor<T, D>` it can still be passed to a function which just needs a
/// store accessor.
///
/// Acquiring an [`Accessor`] can be done through
/// [`StoreContextMut::run_concurrent`] for example or in a host function
/// through
/// [`Linker::func_wrap_concurrent`](crate::component::LinkerInstance::func_wrap_concurrent).
pub trait AsAccessor {
    /// The `T` in `Store<T>` that this accessor refers to.
    type Data: 'static;

    /// The `D` in `Accessor<T, D>`, or the projection out of
    /// `Self::Data`.
    type AccessorData: HasData + ?Sized;

    /// Returns the accessor that this is referring to.
    fn as_accessor(&self) -> &Accessor<Self::Data, Self::AccessorData>;
}

impl<T: AsAccessor + ?Sized> AsAccessor for &T {
    type Data = T::Data;
    type AccessorData = T::AccessorData;

    fn as_accessor(&self) -> &Accessor<Self::Data, Self::AccessorData> {
        T::as_accessor(self)
    }
}

impl<T, D: HasData + ?Sized> AsAccessor for Accessor<T, D> {
    type Data = T;
    type AccessorData = D;

    fn as_accessor(&self) -> &Accessor<T, D> {
        self
    }
}

// Note that it is intentional at this time that `Accessor` does not actually
// store `&mut T` or anything similar. This distinctly enables the `Accessor`
// structure to be both `Send` and `Sync` regardless of what `T` is (or `D` for
// that matter). This is used to ergonomically simplify bindings where the
// majority of the time `Accessor` is closed over in a future which then needs
// to be `Send` and `Sync`. To avoid needing to write `T: Send` everywhere (as
// you already have to write `T: 'static`...) it helps to avoid this.
//
// Note as well that `Accessor` doesn't actually store its data at all. Instead
// it's more of a "proof" of what can be accessed from TLS. API design around
// `Accessor` and functions like `Linker::func_wrap_concurrent` are
// intentionally made to ensure that `Accessor` is ideally only used in the
// context that TLS variables are actually set. For example host functions are
// given `&Accessor`, not `Accessor`, and this prevents them from persisting
// the value outside of a future. Within the future the TLS variables are all
// guaranteed to be set while the future is being polled.
//
// Finally though this is not an ironclad guarantee, but nor does it need to be.
// The TLS APIs are designed to panic or otherwise model usage where they're
// called recursively or similar. It's hoped that code cannot be constructed to
// actually hit this at runtime but this is not a safety requirement at this
// time.
const _: () = {
    const fn assert<T: Send + Sync>() {}
    assert::<Accessor<UnsafeCell<u32>>>();
};

impl<T> Accessor<T> {
    /// Creates a new `Accessor` backed by the specified functions.
    ///
    /// - `get`: used to retrieve the store
    ///
    /// - `get_data`: used to "project" from the store's associated data to
    /// another type (e.g. a field of that data or a wrapper around it).
    ///
    /// - `spawn`: used to queue spawned background tasks to be run later
    pub(crate) fn new(token: StoreToken<T>) -> Self {
        Self {
            token,
            get_data: |x| x,
        }
    }
}

impl<T, D> Accessor<T, D>
where
    D: HasData + ?Sized,
{
    /// Run the specified closure, passing it mutable access to the store.
    ///
    /// This function is one of the main building blocks of the [`Accessor`]
    /// type. This yields synchronous, blocking, access to the store via an
    /// [`Access`]. The [`Access`] implements [`AsContextMut`] in addition to
    /// providing the ability to access `D` via [`Access::get`]. Note that the
    /// `fun` here is given only temporary access to the store and `T`/`D`
    /// meaning that the return value `R` here is not allowed to capture borrows
    /// into the two. If access is needed to data within `T` or `D` outside of
    /// this closure then it must be `clone`d out, for example.
    ///
    /// # Panics
    ///
    /// This function will panic if it is call recursively with any other
    /// accessor already in scope. For example if `with` is called within `fun`,
    /// then this function will panic. It is up to the embedder to ensure that
    /// this does not happen.
    pub fn with<R>(&self, fun: impl FnOnce(Access<'_, T, D>) -> R) -> R {
        tls::get(|vmstore| {
            fun(Access {
                store: self.token.as_context_mut(vmstore),
                get_data: self.get_data,
            })
        })
    }

    /// Returns the getter this accessor is using to project from `T` into
    /// `D::Data`.
    pub fn getter(&self) -> fn(&mut T) -> D::Data<'_> {
        self.get_data
    }

    /// Changes this accessor to access `D2` instead of the current type
    /// parameter `D`.
    ///
    /// This changes the underlying data access from `T` to `D2::Data<'_>`.
    ///
    /// # Panics
    ///
    /// When using this API the returned value is disconnected from `&self` and
    /// the lifetime binding the `self` argument. An `Accessor` only works
    /// within the context of the closure or async closure that it was
    /// originally given to, however. This means that due to the fact that the
    /// returned value has no lifetime connection it's possible to use the
    /// accessor outside of `&self`, the original accessor, and panic.
    ///
    /// The returned value should only be used within the scope of the original
    /// `Accessor` that `self` refers to.
    pub fn with_getter<D2: HasData>(
        &self,
        get_data: fn(&mut T) -> D2::Data<'_>,
    ) -> Accessor<T, D2> {
        Accessor {
            token: self.token,
            get_data,
        }
    }

    /// Spawn a background task which will receive an `&Accessor<T, D>` and
    /// run concurrently with any other tasks in progress for the current
    /// store.
    ///
    /// This is particularly useful for host functions which return a `stream`
    /// or `future` such that the code to write to the write end of that
    /// `stream` or `future` must run after the function returns.
    ///
    /// The returned [`JoinHandle`] may be used to cancel the task.
    ///
    /// # Panics
    ///
    /// Panics if called within a closure provided to the [`Accessor::with`]
    /// function. This can only be called outside an active invocation of
    /// [`Accessor::with`].
    pub fn spawn(&self, task: impl for<'fut> AccessorTask<'fut, T, D>) -> Result<JoinHandle>
    where
        T: 'static,
    {
        let accessor = self.clone_for_spawn();
        self.with(|mut access| access.as_context_mut().spawn_with_accessor(accessor, task))
    }

    fn clone_for_spawn(&self) -> Self {
        Self {
            token: self.token,
            get_data: self.get_data,
        }
    }

    /// Polls to see if this store contains any "interesting" tasks still within
    /// it.
    ///
    /// Returns `Poll::Ready(())` if there are no more interesting tasks, and
    /// otherwise returns `Poll::Pending`. If pending is returned then whenever
    /// the last remaining "interesting" task has exited the provided context's
    /// waker will be notified. Note that only the waker passed to the last call
    /// to `poll_no_interesting_tasks` for the store will be notified, so this
    /// is only appropriate to use once-at-a-time per store.
    ///
    /// The component model specification, as of this current date, does not
    /// have a distinction between "interesting" tasks and not. The current
    /// intention is that in a future revision of the component model this will
    /// be distinguished at the component ABI level where tasks will be able to
    /// flag themselves as "interesting" optionally. Additionally extra work can
    /// be opted-in to being "interesting".
    ///
    /// For now what this means is that all component model tasks within this
    /// store are considered interesting. This specifically includes the entire
    /// duration of a task, so even all of the time after a task has returned
    /// but before it has exited. This means that this function is, today,
    /// effectively a proxy for "are there any more tasks still running in this
    /// store". This can be used by embedders to determine whether there's any
    /// more work going on, even in the background, for a particular guest.
    /// Hosts can use this as a signal that the guest wants to stay alive a
    /// little longer, even after a task has returned.
    ///
    /// In the future this predicate won't include all tasks in this store. Some
    /// tasks will be able to flag themselves as not interesting, meaning that
    /// when this returns ready it'd be possible that there are still tasks
    /// remaining in the store.
    ///
    /// Note that at this time spawned threads within a task are always
    /// considered uninteresting. If this function returns ready, then spawned
    /// threads may still be in the store.
    pub fn poll_no_interesting_tasks(&self, cx: &mut Context<'_>) -> Poll<()> {
        self.with(|mut access| {
            let store = access.as_context_mut().0;
            let state = store.concurrent_state_mut_without_forcing_current_thread();
            if state.interesting_tasks == 0 {
                Poll::Ready(())
            } else {
                state.interesting_tasks_empty_waker = Some(cx.waker().clone());
                Poll::Pending
            }
        })
    }

    /// Poll to see if the component instance corresponding to the specified
    /// function is ready to run a concurrent call without queuing it (i.e. does
    /// not have backpressure enabled and does not have a sync call in
    /// progress).
    ///
    /// Returns `Poll::Ready(())` if the component instance is ready to run a
    /// concurrent call, and otherwise returns `Poll::Pending`.  If pending is
    /// returned then whenever the instance becomes ready for a call the
    /// provided context's waker will be notified.  Note that only the waker
    /// passed to the last call to `poll_ready_for_concurrent_call` for the
    /// store will be notified (regardless of whether the same or different
    /// `Func` is specified relative to earlier calls), so this is only
    /// appropriate to use once-at-a-time per store.  Also note that the waker
    /// may be notified when _any_ instance becomes callable (i.e. not
    /// necessarily the last one polled), so this function must be called again
    /// to determine if the instance of interest is ready.
    pub fn poll_ready_for_concurrent_call(&self, func: Func, cx: &mut Context<'_>) -> Poll<()> {
        self.with(|mut access| {
            let store = access.as_context_mut().0;
            let (_, _, _, raw_options) = func.abi_info(store);
            let instance = func.instance().runtime_instance(raw_options.instance);
            let state = store.instance_state(instance).concurrent_state();
            if state.backpressure == 0 {
                Poll::Ready(())
            } else {
                store
                    .concurrent_state_mut_without_forcing_current_thread()
                    .ready_for_concurrent_call_waker = Some(cx.waker().clone());
                Poll::Pending
            }
        })
    }
}

/// Represents an async closure which may be provided to `Accessor::spawn`,
/// `Accessor::forward`, or `StoreContextMut::spawn`.
// TODO: Replace this with `core::ops::AsyncFnOnce` when we are able to put `Send`
// bound on the unnamed `Future` directly.
//
// As of this writing, it's not possible to specify e.g. `Send` and `Sync`
// bounds on the `Future` type returned by an `AsyncFnOnce`.  Also, using `F:
// Future<Output = Result<()>> + Send + Sync, FN: FnOnce(&Accessor<T>) -> F +
// Send + Sync + 'static` fails with a type mismatch error as we cannot describe
// that `F` should have `&Accessor<T>`'s unnamed lifetime
//
// Instead, this trait is used as a workaround for this limitation, the bound on `Self`
// implementing `AsyncFnOnce()` is required for Rust to automatically infer that an async
// closure is to be provided wherever we are accepting `impl for<'fut> AccessorTask<'fut, T, D>`
// as argument. Otherwise, users will have to fully qualify the closure types before it
// is accepted as a valid value. This also means that it is not intended for a user
// to manually implement this trait for any arbitrary type, since they would first
// have to implement `AsyncFnOnce`, which is unstable.
//
// The blanket implementation for this trait will ensure that `Self` is an async closure
// that returns a `Future` that is `Send`, and lives as long as `&Accessor<T, D>`
pub trait AccessorTask<'fut, T, D = HasSelf<T>>:
    AsyncFnOnce(&Accessor<T, D>) -> Result<()> + Send + 'static
where
    D: HasData + ?Sized,
{
    /// Run the task.
    fn run(self, accessor: &'fut Accessor<T, D>) -> impl Future<Output = Result<()>> + Send + 'fut;
}

impl<'fut, F, Fut, T, D> AccessorTask<'fut, T, D> for F
where
    T: 'static,
    F: AsyncFnOnce(&Accessor<T, D>) -> Result<()>,
    F: FnOnce(&'fut Accessor<T, D>) -> Fut + Send + 'static,
    Fut: Future<Output = Result<()>> + Send + 'fut,
    D: HasData,
{
    fn run(self, accessor: &'fut Accessor<T, D>) -> impl Future<Output = Result<()>> + Send + 'fut {
        (self)(accessor)
    }
}

/// Represents parameter and result metadata for the caller side of a
/// guest->guest call orchestrated by a fused adapter.
enum CallerInfo {
    /// Metadata for a call to an async-lowered import
    Async {
        params: Vec<ValRaw>,
        has_result: bool,
    },
    /// Metadata for a call to an sync-lowered import
    Sync {
        params: Vec<ValRaw>,
        result_count: u32,
    },
}

/// Indicates how a guest task is waiting on a waitable set.
enum WaitMode {
    /// The guest task is waiting using `task.wait`
    Fiber(StoreFiber<'static>),
    /// The guest task is waiting via a callback declared as part of an
    /// async-lifted export.
    Callback(Instance),
}

impl fmt::Debug for WaitMode {
    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
        match self {
            Self::Fiber(_) => f.debug_tuple("Fiber").finish(),
            Self::Callback(instance) => f.debug_tuple("Callback").field(instance).finish(),
        }
    }
}

/// Represents the reason a fiber is suspending itself.
#[derive(Debug)]
enum SuspendReason {
    /// The fiber is waiting for an event to be delivered to the specified
    /// waitable set or task.
    Waiting {
        set: TableId<WaitableSet>,
        thread: QualifiedThreadId,
    },
    /// The fiber is waiting for a subtask to suspend or exit, e.g. for a
    /// guest-to-guest call or a `subtask.cancel`.
    YieldingToSubtask { thread: QualifiedThreadId },
    /// The fiber has finished handling its most recent work item and is waiting
    /// for another (or to be dropped if it is no longer needed).
    NeedWork,
    /// The fiber is yielding and should be resumed once other tasks have had a
    /// chance to run.
    Yielding { thread: QualifiedThreadId },
    /// The fiber was explicitly suspended with a call to `thread.suspend` or
    /// `thread.switch-to`.
    ExplicitlySuspending { thread: QualifiedThreadId },
}

/// Represents a pending call into guest code for a given guest task.
enum GuestCallKind {
    /// Indicates there's an event to deliver to the task, possibly related to a
    /// waitable set the task has been waiting on or polling.
    DeliverEvent {
        /// The instance to which the task belongs.
        instance: Instance,
        /// The waitable set the event belongs to, if any.
        ///
        /// If this is `None` the event will be waiting in the
        /// `GuestTask::event` field for the task.
        set: Option<TableId<WaitableSet>>,
    },
    /// Indicates that a new guest task call is pending and may be executed
    /// using the specified closure.
    ///
    /// If the closure returns `Ok(Some(call))`, the `call` should be run
    /// immediately using `handle_guest_call`.
    StartImplicit(Box<dyn FnOnce(&mut dyn VMStore) -> Result<()> + Send + Sync>),
    StartExplicit(Box<dyn FnOnce(&mut dyn VMStore) -> Result<()> + Send + Sync>),
}

impl fmt::Debug for GuestCallKind {
    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
        match self {
            Self::DeliverEvent { instance, set } => f
                .debug_struct("DeliverEvent")
                .field("instance", instance)
                .field("set", set)
                .finish(),
            Self::StartImplicit(_) => f.debug_tuple("StartImplicit").finish(),
            Self::StartExplicit(_) => f.debug_tuple("StartExplicit").finish(),
        }
    }
}

/// The target of a suspension intrinsic.
#[derive(Copy, Clone, Debug)]
pub enum SuspensionTarget {
    Resume(u32),
    Promote(u32),
    None,
}

/// Behavior for `resume_thread`.
#[derive(Copy, Clone, Debug)]
pub enum ResumeThread {
    Promote,
    Resume,
    ResumeLater,
}

/// Represents a pending call into guest code for a given guest thread.
#[derive(Debug)]
struct GuestCall {
    thread: QualifiedThreadId,
    kind: GuestCallKind,
}

impl GuestCall {
    /// Returns whether or not the call is ready to run.
    ///
    /// A call will not be ready to run if either:
    ///
    /// - the (sub-)component instance to be called has already been entered and
    /// cannot be reentered until an in-progress call completes
    ///
    /// - the call is for a not-yet started task and the (sub-)component
    /// instance to be called has backpressure enabled
    fn is_ready(&self, store: &mut StoreOpaque) -> Result<bool> {
        let task = store.concurrent_state_mut()?.get_mut(self.thread.task)?;
        let async_typed = task.async_typed;
        let instance = task.instance;
        let state = store.instance_state(instance).concurrent_state();

        let ready = match &self.kind {
            GuestCallKind::DeliverEvent { .. } => !state.do_not_enter,
            GuestCallKind::StartImplicit(_) => {
                !async_typed || !(state.do_not_enter || state.backpressure > 0)
            }
            GuestCallKind::StartExplicit(_) => true,
        };
        log::trace!(
            "call {self:?} ready? {ready} (do_not_enter: {}; backpressure: {})",
            state.do_not_enter,
            state.backpressure
        );
        Ok(ready)
    }
}

/// Job to be run on a worker fiber.
enum WorkerItem {
    GuestCall(GuestCall),
    Function(AlwaysMut<Box<dyn FnOnce(&mut dyn VMStore) -> Result<()> + Send>>),
}

/// Represents a pending work item to be handled by the event loop for a given
/// component instance.
enum WorkItem {
    /// A host task to be pushed to `ConcurrentState::futures`.
    PushFuture(AlwaysMut<HostTaskFuture>),
    /// A fiber to resume.
    ResumeFiber {
        instance: RuntimeInstance,
        thread: QualifiedThreadId,
        fiber: StoreFiber<'static>,
    },
    /// A thread to resume.
    ResumeThread {
        instance: RuntimeInstance,
        thread: QualifiedThreadId,
    },
    /// A pending call into guest code for a given guest task.
    GuestCall {
        instance: RuntimeInstance,
        call: GuestCall,
    },
    /// A job to run on a worker fiber.
    WorkerFunction(AlwaysMut<Box<dyn FnOnce(&mut dyn VMStore) -> Result<()> + Send>>),
}

/// A fiber temporarily outside the store's scheduler state.
///
/// Construct this before returning a future so cancellation before its first
/// poll also disposes of the fiber. Successful handoffs take the fiber only
/// after all fallible destination checks have completed.
struct ResumingFiber<'a> {
    store: &'a mut StoreOpaque,
    fiber: Option<StoreFiber<'static>>,
}

impl Drop for ResumingFiber<'_> {
    fn drop(&mut self) {
        if let Some(mut fiber) = self.fiber.take() {
            fiber.dispose(self.store);
        }
    }
}

impl fmt::Debug for WorkItem {
    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
        match self {
            Self::PushFuture(_) => f.debug_tuple("PushFuture").finish(),
            Self::ResumeFiber {
                instance, thread, ..
            } => f
                .debug_struct("ResumeFiber")
                .field("instance", instance)
                .field("thread", thread)
                .finish(),
            Self::ResumeThread { instance, thread } => f
                .debug_struct("ResumeThread")
                .field("instance", instance)
                .field("thread", thread)
                .finish(),
            Self::GuestCall { instance, call } => f
                .debug_struct("GuestCall")
                .field("instance", instance)
                .field("call", call)
                .finish(),
            Self::WorkerFunction(_) => f.debug_tuple("WorkerFunction").finish(),
        }
    }
}

/// Whether a suspension intrinsic was cancelled or completed
#[derive(Copy, Clone, Ord, PartialOrd, Eq, PartialEq)]
pub(crate) enum WaitResult {
    Cancelled,
    Completed,
}

/// Poll the specified future until it completes on behalf of a guest->host call
/// using a sync-lowered import.
///
/// This is similar to `Instance::first_poll` except it's for sync-lowered
/// imports, meaning we don't need to handle cancellation and we can block the
/// caller until the task completes, at which point the caller can handle
/// lowering the result to the guest's stack and linear memory.
pub(crate) fn poll_and_block<R: Send + Sync + 'static>(
    store: &mut dyn VMStore,
    host_task: EnteredHostTask,
    future: impl Future<Output = Result<R>> + Send + 'static,
) -> Result<R> {
    // Poll the future once before creating a host task. The host task will be
    // created lazily if it's needed during the poll and otherwise will be
    // created if the future suspends.  We can use a dummy `Waker` here because
    // we'll add the future to `ConcurrentState::futures` and poll it
    // automatically from the event loop if it doesn't complete immediately
    // here.
    let mut future = Box::pin(future);
    let poll = tls::set(store, || {
        future
            .as_mut()
            .poll(&mut Context::from_waker(&Waker::noop()))
    });

    let caller = match host_task {
        Some(caller) => caller,
        None => bail_bug!("host task wasn't created but should have been"),
    };

    let task = match poll {
        // It completed immediately, so no persistent host task is needed.
        Poll::Ready(result) => return result,

        // It did not complete immediately; create the host task and add it to
        // `ConcurrentState::futures` so it will be polled via the event loop;
        // then use `GuestThread::sync_call_set` to wait for the task to
        // complete, suspending the current fiber until it does so.
        Poll::Pending => {
            let Some(task) = store.materialize_host_task_id()? else {
                bail_bug!("current thread is not a host thread")
            };

            // Wrap the future in a closure which will stash its result in the
            // host task and resume this fiber when it completes.
            let future = Box::pin(async move {
                let result = run_with_host_task_set(task, future).await??;
                tls::get(move |store| {
                    let state = store.concurrent_state_mut()?;
                    let host_state = &mut state.get_mut(task)?.state;
                    assert!(matches!(host_state, HostTaskState::CalleeStarted));
                    *host_state = HostTaskState::CalleeFinished(Box::new(result));

                    Waitable::Host(task).set_event(
                        state,
                        Some(Event::Subtask {
                            status: Status::Returned,
                        }),
                    )?;

                    Ok(())
                })
            }) as HostTaskFuture;

            let caller_instance = store.concurrent_state_mut()?.get_mut(caller.task)?.instance;
            store.switch_or_trap_if_may_not_suspend(caller_instance)?;

            let state = store.concurrent_state_mut()?;
            state.push_future(future);

            let set = state.get_mut(caller.thread)?.sync_call_set;
            Waitable::Host(task).join(state, Some(set))?;

            store.suspend(SuspendReason::Waiting {
                set,
                thread: caller,
            })?;

            // Remove the `task` from the `sync_call_set` to ensure that when
            // this function returns and the task is deleted that there are no
            // more lingering references to this host task.
            Waitable::Host(task).join(store.concurrent_state_mut()?, None)?;
            task
        }
    };

    // Retrieve and return the result.
    let host_state = &mut store.concurrent_state_mut()?.get_mut(task)?.state;
    match mem::replace(host_state, HostTaskState::CalleeDone { cancelled: false }) {
        HostTaskState::CalleeFinished(result) => Ok(match result.downcast() {
            Ok(result) => *result,
            Err(_) => bail_bug!("host task finished with wrong type of result"),
        }),
        _ => bail_bug!("unexpected host task state after completion"),
    }
}

/// Execute the specified guest call.
fn handle_guest_call(store: &mut dyn VMStore, call: GuestCall) -> Result<()> {
    match call.kind {
        GuestCallKind::DeliverEvent { instance, set } => {
            // This thread's wait on `set`, if any, which started in
            // `Instance::wait_with_callback`, ends here.
            if let Some(set) = set {
                store.concurrent_state_mut()?.get_mut(set)?.stop_waiting()?;
            }
            let (event, waitable) = match instance.get_event(store, call.thread.task, set, true)? {
                Some(pair) => pair,
                None => match set {
                    // The event which woke this thread up was taken by another
                    // thread (e.g. via `waitable-set.poll`) or its waitable was
                    // removed from the set before this thread had a chance to
                    // run. Regardless, keep waiting.
                    Some(set) => {
                        log::trace!(
                            "event for {:?} on {set:?} no longer present; waiting again",
                            call.thread
                        );
                        return instance.wait_with_callback(
                            store.store_opaque_mut(),
                            call.thread,
                            set,
                        );
                    }
                    // When a thread yields it doesn't wait on a set and the
                    // event to be delivered isn't stored anywhere, and this is
                    // just handling that case so indicate that the yield is
                    // complete.
                    None => (Event::None, None),
                },
            };
            let state = store.concurrent_state_mut()?;
            // If this thread was woken up by an event in `set` but is instead
            // receiving its own task's event (e.g. a cancellation), then the
            // set's event is left for another thread waiting on it, so wake
            // the next waiter up if there is one.
            if let Some(set) = set
                && waitable.is_none()
                && !state.get_mut(set)?.ready.is_empty()
            {
                state.wake_waiter(set)?;
            }
            let task = state.get_mut(call.thread.task)?;
            let runtime_instance = task.instance;
            let handle = waitable.map(|(_, v)| v).unwrap_or(0);

            log::trace!(
                "use callback to deliver event {event:?} to {:?} for {waitable:?}",
                call.thread,
            );

            let old_thread = store.set_thread(call.thread)?;
            log::trace!(
                "GuestCallKind::DeliverEvent: replaced {old_thread:?} with {:?} as current thread",
                call.thread
            );

            store.enter_instance(runtime_instance);

            let Some(callback) = store
                .concurrent_state_mut()?
                .get_mut(call.thread.task)?
                .callback
                .take()
            else {
                bail_bug!("guest task callback field not present")
            };

            let code = callback(store, event, handle)?;

            store
                .concurrent_state_mut()?
                .get_mut(call.thread.task)?
                .callback = Some(callback);

            store.exit_instance(runtime_instance)?;

            store.set_thread(old_thread)?;

            instance.handle_callback_code(store, call.thread, runtime_instance.index, code)?;

            log::trace!("GuestCallKind::DeliverEvent: restored {old_thread:?} as current thread");
        }
        GuestCallKind::StartImplicit(fun) => {
            fun(store)?;
        }
        GuestCallKind::StartExplicit(fun) => {
            fun(store)?;
        }
    }

    Ok(())
}

impl<T> Store<T> {
    /// Convenience wrapper for [`StoreContextMut::run_concurrent`].
    pub async fn run_concurrent<R>(&mut self, fun: impl AsyncFnOnce(&Accessor<T>) -> R) -> Result<R>
    where
        T: Send + 'static,
    {
        ensure!(
            self.as_context().0.concurrency_support(),
            "cannot use `run_concurrent` when Config::concurrency_support disabled",
        );
        self.as_context_mut().run_concurrent(fun).await
    }

    #[doc(hidden)]
    pub fn assert_concurrent_state_empty(&mut self) {
        self.as_context_mut().assert_concurrent_state_empty();
    }

    #[doc(hidden)]
    pub fn concurrent_state_table_size(&mut self) -> usize {
        self.as_context_mut().concurrent_state_table_size()
    }

    /// Convenience wrapper for [`StoreContextMut::spawn`].
    pub fn spawn(
        &mut self,
        task: impl for<'fut> AccessorTask<'fut, T, HasSelf<T>>,
    ) -> Result<JoinHandle>
    where
        T: 'static,
    {
        self.as_context_mut().spawn(task)
    }
}

impl<T> StoreContextMut<'_, T> {
    /// Assert that all the relevant tables and queues in the concurrent state
    /// for this store are empty.
    ///
    /// This is for sanity checking in integration tests
    /// (e.g. `component-async-tests`) that the relevant state has been cleared
    /// after each test concludes.  This should help us catch leaks, e.g. guest
    /// tasks which haven't been deleted despite having completed and having
    /// been dropped by their supertasks.
    ///
    /// Only intended for use in Wasmtime's own testing.
    #[doc(hidden)]
    pub fn assert_concurrent_state_empty(self) {
        let store = self.0;
        store
            .store_data_mut()
            .components
            .assert_instance_states_empty();
        let state = store.concurrent_state_mut().unwrap();
        assert!(
            state.table.get_mut().is_empty(),
            "non-empty table: {:?}",
            state.table.get_mut()
        );
        assert!(state.switch_item.is_none());
        assert!(state.next_switch_item.is_none());
        assert!(state.high_priority.is_empty());
        assert!(state.low_priority.is_empty());
        assert!(state.saved_next_switch_items.is_empty());
        assert!(state.unforced_current_thread.is_none());
        assert!(state.deferred_host_call_context.is_none());
        assert!(state.futures_mut().unwrap().is_empty());
        assert!(state.global_error_context_ref_counts.is_empty());
    }

    /// Helper function to perform tests over the size of the concurrent state
    /// table which can be useful for detecting leaks.
    ///
    /// Only intended for use in Wasmtime's own testing.
    #[doc(hidden)]
    pub fn concurrent_state_table_size(&mut self) -> usize {
        self.0
            .concurrent_state_mut()
            .unwrap()
            .table
            .get_mut()
            .iter_mut()
            .count()
    }

    /// Spawn a background task to run as part of this instance's event loop.
    ///
    /// The task will receive an `&Accessor<U>` and run concurrently with
    /// any other tasks in progress for the instance.
    ///
    /// Note that the task will only make progress if and when the event loop
    /// for this instance is run.
    ///
    /// The returned [`JoinHandle`] may be used to cancel the task.
    pub fn spawn(mut self, task: impl for<'fut> AccessorTask<'fut, T>) -> Result<JoinHandle>
    where
        T: 'static,
    {
        let accessor = Accessor::new(StoreToken::new(self.as_context_mut()));
        self.spawn_with_accessor(accessor, task)
    }

    /// Internal implementation of `spawn` functions where a `store` is
    /// available along with an `Accessor`.
    fn spawn_with_accessor<D>(
        self,
        accessor: Accessor<T, D>,
        task: impl for<'fut> AccessorTask<'fut, T, D>,
    ) -> Result<JoinHandle>
    where
        T: 'static,
        D: HasData + ?Sized,
    {
        // Create an "abortable future" here where internally the future will
        // hook calls to poll and possibly spawn more background tasks on each
        // iteration.
        let (handle, future) = JoinHandle::run(async move { task.run(&accessor).await });
        self.0
            .concurrent_state_mut()?
            .push_future(Box::pin(async move { future.await.unwrap_or(Ok(())) }));
        Ok(handle)
    }

    /// Run the specified closure `fun` to completion as part of this store's
    /// event loop.
    ///
    /// This will run `fun` as part of this store's event loop until it
    /// yields a result.  `fun` is provided an [`Accessor`], which provides
    /// controlled access to the store and its data.
    ///
    /// This function can be used to invoke [`Func::call_concurrent`] for
    /// example within the async closure provided here.
    ///
    /// This function will unconditionally return an error if
    /// [`Config::concurrency_support`] is disabled.
    ///
    /// [`Config::concurrency_support`]: crate::Config::concurrency_support
    ///
    /// # Store-blocking behavior
    ///
    /// At this time there are certain situations in which the `Future` returned
    /// by the `AsyncFnOnce` passed to this function will not be polled for an
    /// extended period of time, despite one or more `Waker::wake` events having
    /// occurred for the task to which it belongs.  This can manifest as the
    /// `Future` seeming to be "blocked" or "locked up", but is actually due to
    /// the `Store` being held by e.g. a blocking host function, preventing the
    /// `Future` from being polled. A canonical example of this is when the
    /// `fun` provided to this function attempts to set a timeout for an
    /// invocation of a wasm function. In this situation the async closure is
    /// waiting both on (a) the wasm computation to finish, and (b) the timeout
    /// to elapse. At this time this setup will not always work and the timeout
    /// may not reliably fire.
    ///
    /// This function will not block the current thread and as such is always
    /// suitable to run in an `async` context, but the current implementation of
    /// Wasmtime can lead to situations where a certain wasm computation is
    /// required to make progress the closure to make progress. This is an
    /// artifact of Wasmtime's historical implementation of `async` functions
    /// and is the topic of [#11869] and [#11870]. In the timeout example from
    /// above it means that Wasmtime can get "wedged" for a bit where (a) must
    /// progress for a readiness notification of (b) to get delivered.
    ///
    /// This effectively means that it's not possible to reliably perform a
    /// "select" operation within the `fun` closure, which timeouts for example
    /// are based on. Fixing this requires some relatively major refactoring
    /// work within Wasmtime itself. This is a known pitfall otherwise and one
    /// that is intended to be fixed one day. In the meantime it's recommended
    /// to apply timeouts or such to the entire `run_concurrent` call itself
    /// rather than internally.
    ///
    /// [#11869]: https://github.com/bytecodealliance/wasmtime/issues/11869
    /// [#11870]: https://github.com/bytecodealliance/wasmtime/issues/11870
    ///
    /// # Example
    ///
    /// ```
    /// # use {
    /// #   wasmtime::{
    /// #     error::{Result},
    /// #     component::{ Component, Linker, Resource, ResourceTable},
    /// #     Config, Engine, Store
    /// #   },
    /// # };
    /// #
    /// # struct MyResource(u32);
    /// # struct Ctx { table: ResourceTable }
    /// #
    /// # async fn foo() -> Result<()> {
    /// # let mut config = Config::new();
    /// # let engine = Engine::new(&config)?;
    /// # let mut store = Store::new(&engine, Ctx { table: ResourceTable::new() });
    /// # let mut linker = Linker::new(&engine);
    /// # let component = Component::new(&engine, "")?;
    /// # let instance = linker.instantiate_async(&mut store, &component).await?;
    /// # let foo = instance.get_typed_func::<(Resource<MyResource>,), (Resource<MyResource>,)>(&mut store, "foo")?;
    /// # let bar = instance.get_typed_func::<(u32,), ()>(&mut store, "bar")?;
    /// store.run_concurrent(async |accessor| -> wasmtime::Result<_> {
    ///    let resource = accessor.with(|mut access| access.get().table.push(MyResource(42)))?;
    ///    let (another_resource,) = foo.call_concurrent(accessor, (resource,)).await?;
    ///    let value = accessor.with(|mut access| access.get().table.delete(another_resource))?;
    ///    bar.call_concurrent(accessor, (value.0,)).await?;
    ///    Ok(())
    /// }).await??;
    /// # Ok(())
    /// # }
    /// ```
    pub async fn run_concurrent<R>(self, fun: impl AsyncFnOnce(&Accessor<T>) -> R) -> Result<R>
    where
        T: Send + 'static,
    {
        ensure!(
            self.0.concurrency_support(),
            "cannot use `run_concurrent` when Config::concurrency_support disabled",
        );
        self.do_run_concurrent(fun, false).await
    }

    pub(super) async fn run_concurrent_trap_on_idle<R>(
        self,
        fun: impl AsyncFnOnce(&Accessor<T>) -> R,
    ) -> Result<R> {
        self.do_run_concurrent(fun, true).await
    }

    async fn do_run_concurrent<R>(
        mut self,
        fun: impl AsyncFnOnce(&Accessor<T>) -> R,
        trap_on_idle: bool,
    ) -> Result<R> {
        debug_assert!(self.0.concurrency_support());
        let already_running = self
            .0
            .concurrent_state_mut_already_forced_current_thread()
            .event_loop_running;
        if already_running {
            bail!("Recursive `StoreContextMut::run_concurrent` calls not supported")
        }
        let token = StoreToken::new(self.as_context_mut());

        struct Dropper<'a, T: 'static, V> {
            store: StoreContextMut<'a, T>,
            value: ManuallyDrop<V>,
        }

        impl<'a, T, V> Drop for Dropper<'a, T, V> {
            fn drop(&mut self) {
                self.store
                    .0
                    .concurrent_state_mut_already_forced_current_thread()
                    .event_loop_running = false;

                tls::set(self.store.0, || {
                    // SAFETY: Here we drop the value without moving it for the
                    // first and only time -- per the contract for `Drop::drop`,
                    // this code won't run again, and the `value` field will no
                    // longer be accessible.
                    unsafe { ManuallyDrop::drop(&mut self.value) }
                });
            }
        }

        let accessor = &Accessor::new(token);
        self.0
            .concurrent_state_mut_already_forced_current_thread()
            .event_loop_running = true;
        let dropper = &mut Dropper {
            store: self,
            value: ManuallyDrop::new(fun(accessor)),
        };
        // SAFETY: We never move `dropper` nor its `value` field.
        let future = unsafe { Pin::new_unchecked(dropper.value.deref_mut()) };

        let result = dropper
            .store
            .as_context_mut()
            .poll_until(future, trap_on_idle)
            .await;

        if result.is_err() {
            dropper.store.0.set_trapped();
        }

        result
    }

    /// Run this store's event loop.
    ///
    /// The returned future will resolve when the specified future completes or,
    /// if `trap_on_idle` is true, when the event loop can't make further
    /// progress.
    async fn poll_until<R>(
        mut self,
        mut future: Pin<&mut impl Future<Output = R>>,
        trap_on_idle: bool,
    ) -> Result<R> {
        struct Reset<'a, T: 'static> {
            store: StoreContextMut<'a, T>,
            futures: Option<FuturesUnordered<HostTaskFuture>>,
        }

        impl<'a, T> Drop for Reset<'a, T> {
            fn drop(&mut self) {
                if let Some(futures) = self.futures.take() {
                    *self
                        .store
                        .0
                        .concurrent_state_mut_already_forced_current_thread()
                        .futures
                        .get_mut() = Some(futures);
                }
            }
        }

        // Keep a continuously busy event loop from holding one executor poll
        // indefinitely.  In particular, Tokio's cooperative budget is only
        // replenished when the future returns `Pending` to the executor.
        const MAX_TURNS_WITHOUT_YIELD: usize = 128;
        let mut turns_without_yield = 0;

        loop {
            // Take `ConcurrentState::futures` out of the store so we can poll
            // it while also safely giving any of the futures inside access to
            // `self`.
            let futures = self.0.concurrent_state_mut()?.futures.get_mut().take();
            let mut reset = Reset {
                store: self.as_context_mut(),
                futures,
            };
            let mut next = match reset.futures.as_mut() {
                Some(f) => pin!(f.next()),
                None => bail_bug!("concurrent state missing futures field"),
            };

            enum PollResult<R> {
                Complete(R),
                ProcessWork {
                    ready: Option<WorkItem>,
                    low_priority: bool,
                },
            }

            let result = future::poll_fn(|cx| {
                // First, poll the future we were passed as an argument and
                // return immediately if it's ready.
                if let Poll::Ready(value) = tls::set(reset.store.0, || future.as_mut().poll(cx)) {
                    return Poll::Ready(Ok(PollResult::Complete(value)));
                }

                // If the store has been poisoned by a trap, either before this
                // event loop started or while it was running, then refuse to
                // make any further progress. Any work items, suspended fibers,
                // or pending futures left over from the trap might otherwise
                // resume guest code or observe scheduler state which was never
                // unwound. Note that this is checked after polling `future`
                // above so that it can observe the original error, e.g. a
                // trap delivered over a oneshot channel by a work item.
                if reset.store.0.trapped() {
                    return Poll::Ready(Err(Trap::CannotEnterComponent.into()));
                }

                // Next, poll `ConcurrentState::futures` (which includes any
                // pending host tasks and/or background tasks), returning
                // immediately if one of them fails.
                let next = match tls::set(reset.store.0, || next.as_mut().poll(cx)) {
                    Poll::Ready(Some(output)) => {
                        match output {
                            Err(e) => return Poll::Ready(Err(e)),
                            Ok(()) => {}
                        }
                        Poll::Ready(true)
                    }
                    Poll::Ready(None) => Poll::Ready(false),
                    Poll::Pending => Poll::Pending,
                };

                // Next, identify the next work item to process, if any, using
                // the following priority order:
                //
                // - `switch_item`: Represents the guest thread we _must_ switch
                // to before any other thread runs per the determinism
                // requirements in the Component Model spec.
                //
                // - `high_priority`: "Urgent" work items, e.g. async calls have
                // become freshly unblocked due to backpressure clearing or
                // similar, stream or future state updates, etc.
                //
                // - `low_priority`: Work items such as resuming a fiber after
                // it yields, in which case the point is to let other items run
                // first.
                let state = reset.store.0.concurrent_state_mut()?;
                let mut ready = state.switch_item.take();
                let mut low_priority = false;
                if ready.is_none() {
                    ready = state.high_priority.pop_back();
                    if ready.is_none() {
                        ready = state.low_priority.pop_back();
                        low_priority = true;
                    }
                }
                if ready.is_some() {
                    return Poll::Ready(Ok(PollResult::ProcessWork {
                        ready,
                        low_priority,
                    }));
                }

                // Finally, if we have nothing else to do right now, determine what to do
                // based on whether there are any pending futures in
                // `ConcurrentState::futures`.
                return match next {
                    Poll::Ready(true) => {
                        // In this case, one of the futures in
                        // `ConcurrentState::futures` completed
                        // successfully, so we return now and continue
                        // the outer loop in case there is another one
                        // ready to complete.
                        Poll::Ready(Ok(PollResult::ProcessWork {
                            ready: None,
                            low_priority: false,
                        }))
                    }
                    Poll::Ready(false) => {
                        // Poll the future we were passed one last time
                        // in case one of `ConcurrentState::futures` had
                        // the side effect of unblocking it.
                        if let Poll::Ready(value) =
                            tls::set(reset.store.0, || future.as_mut().poll(cx))
                        {
                            Poll::Ready(Ok(PollResult::Complete(value)))
                        } else {
                            // In this case, there are no more pending
                            // futures in `ConcurrentState::futures`,
                            // there are no remaining work items, _and_
                            // the future we were passed as an argument
                            // still hasn't completed.
                            if trap_on_idle {
                                // `trap_on_idle` is true, so we exit
                                // immediately.

                                // If there are any tasks belonging to
                                // an instance which may not suspend, trap with
                                // `CannotBlockSyncTask`:
                                Poll::Ready(Err(if reset.store.0.any_may_not_suspend()? {
                                    Trap::CannotBlockSyncTask.into()
                                } else {
                                    // Otherwise, trap with `AsyncDeadlock`:
                                    Trap::AsyncDeadlock.into()
                                }))
                            } else {
                                // `trap_on_idle` is false, so we assume
                                // that future will wake up and give us
                                // more work to do when it's ready to.
                                Poll::Pending
                            }
                        }
                    }
                    // There is at least one pending future in
                    // `ConcurrentState::futures` and we have nothing
                    // else to do but wait for now, so we return
                    // `Pending`.
                    Poll::Pending => Poll::Pending,
                };
            })
            .await;

            // Put the `ConcurrentState::futures` back into the store before we
            // return or handle any work items since one or more of those items
            // might append more futures.
            drop(reset);

            match result? {
                // The future we were passed as an argument completed, so we
                // return the result.
                PollResult::Complete(value) => break Ok(value),
                // The future we were passed has not yet completed, so handle
                // any work items and then loop again.
                PollResult::ProcessWork {
                    ready,
                    low_priority,
                } => {
                    struct Dispose<'a, T: 'static> {
                        store: StoreContextMut<'a, T>,
                        ready: Option<WorkItem>,
                    }

                    impl<'a, T> Drop for Dispose<'a, T> {
                        fn drop(&mut self) {
                            if let Some(item) = self.ready.take() {
                                match item {
                                    WorkItem::ResumeFiber { mut fiber, .. } => {
                                        fiber.dispose(self.store.0);
                                    }
                                    WorkItem::PushFuture(future) => {
                                        tls::set(self.store.0, move || drop(future))
                                    }
                                    _ => {}
                                }
                            }
                        }
                    }

                    let mut dispose = Dispose {
                        store: self.as_context_mut(),
                        ready,
                    };

                    // If we're about to run a low-priority task, first yield to
                    // the executor.  This ensures that it won't be starved of
                    // the ability to e.g. update the readiness of sockets,
                    // etc. which the guest may be using `thread.yield` along
                    // with `waitable-set.poll` to monitor in a CPU-heavy loop.
                    //
                    // This works for e.g. `thread.yield` and callbacks which
                    // return `CALLBACK_CODE_YIELD` because we queue a low
                    // priority item to resume the task (i.e. resume the thread
                    // or call the callback, respectively) just prior to
                    // suspending it.  Indeed, as of this writing those are the
                    // _only_ situations we queue low-priority tasks.
                    // Therefore, we interpret the guest's request to yield as
                    // meaning "yield to other guest tasks _and_/_or_ host
                    // operations such as updating socket readiness", the latter
                    // being the async runtime's responsibility.
                    //
                    // In the future, if this ends up causing measurable
                    // performance issues, this could be optimized such that we
                    // only yield periodically (e.g. for batches of low priority
                    // items) and not for each and every individual item.
                    if low_priority {
                        dispose.store.0.yield_now().await;
                        turns_without_yield = 0;
                    }

                    if let Some(item) = dispose.ready.take() {
                        dispose
                            .store
                            .as_context_mut()
                            .handle_work_item(item)
                            .await?;
                    }

                    turns_without_yield += 1;
                    if turns_without_yield == MAX_TURNS_WITHOUT_YIELD {
                        turns_without_yield = 0;
                        dispose.store.0.yield_now().await;
                    }
                }
            }
        }
    }

    /// Handle the specified work item, possibly resuming a fiber if applicable.
    fn handle_work_item(self, item: WorkItem) -> impl Future<Output = Result<()>> {
        // Construct the guarded future before returning, even when the caller
        // never polls this work item. Other variants keep their fibers in the
        // store until their first poll.
        let item = match item {
            WorkItem::ResumeFiber {
                instance,
                thread,
                fiber,
            } => {
                let future = self.0.resume_fiber(fiber);
                log::trace!(
                    "handle work item ResumeFiber {{ instance: {instance:?}, thread: {thread:?} }}"
                );
                return futures::future::Either::Left(future);
            }
            item => item,
        };
        futures::future::Either::Right(async move {
            log::trace!("handle work item {item:?}");
            match item {
                WorkItem::PushFuture(future) => {
                    self.0
                        .concurrent_state_mut()?
                        .futures_mut()?
                        .push(future.into_inner());
                }
                WorkItem::ResumeFiber { .. } => unreachable!(),
                WorkItem::ResumeThread { thread, .. } => {
                    let slot = &mut self.0.concurrent_state_mut()?.get_mut(thread.thread)?.state;
                    if !matches!(slot, GuestThreadState::Ready { .. }) {
                        bail_bug!("cannot resume non-pending thread {thread:?}");
                    }
                    let GuestThreadState::Ready { fiber } =
                        mem::replace(slot, GuestThreadState::Running)
                    else {
                        unreachable!()
                    };
                    self.0.resume_fiber(fiber).await?;
                }
                WorkItem::GuestCall { call, .. } => {
                    if call.is_ready(self.0)? {
                        self.0
                            .concurrent_state_mut()?
                            .get_mut(call.thread.thread)?
                            .wake_on_cancel = WakeOnCancel::None;
                        self.run_on_worker(WorkerItem::GuestCall(call)).await?;
                    } else {
                        let state = self.0.concurrent_state_mut()?;
                        let task = state.get_mut(call.thread.task)?;
                        if !task.starting_sent {
                            task.starting_sent = true;
                            if let GuestCallKind::StartImplicit(_) = &call.kind {
                                Waitable::Guest(call.thread.task).set_event(
                                    state,
                                    Some(Event::Subtask {
                                        status: Status::Starting,
                                    }),
                                )?;
                            }
                        }

                        let instance = state.get_mut(call.thread.task)?.instance;
                        // If this call is for a callback-lifted task waiting on a
                        // waitable set, it can't run until `instance` is enterable
                        // again. Meanwhile, the set may have an event which another
                        // thread waiting on it could take, and that thread may
                        // even be the reason `instance` isn't enterable, so wake
                        // the next waiter up as well if there's still an event
                        // for it. Whichever thread runs first takes the event, and
                        // the other goes back to waiting if none is left.
                        if let GuestCallKind::DeliverEvent { set: Some(set), .. } = &call.kind {
                            let state = self.0.concurrent_state_mut()?;
                            if !state.get_mut(*set)?.ready.is_empty() {
                                state.wake_waiter(*set)?;
                            }
                        }
                        self.0
                            .instance_state(instance)
                            .concurrent_state()
                            .pending
                            .insert(call.thread, call.kind);

                        // Switch back to the caller (or canceller) immediately if
                        // applicable since we aren't yet able to run the subtask it
                        // yielded to.
                        self.0.concurrent_state_mut()?.take_next_switch_item()?;
                    }
                }
                WorkItem::WorkerFunction(fun) => {
                    self.run_on_worker(WorkerItem::Function(fun)).await?;
                }
            }

            Ok(())
        })
    }

    /// Execute the specified guest call on a worker fiber.
    async fn run_on_worker(self, item: WorkerItem) -> Result<()> {
        let state = self.0.concurrent_state_mut()?;
        if state.worker_item.is_some() {
            bail_bug!("worker item already set when taking worker fiber");
        }
        let worker = if let Some(fiber) = state.worker.take() {
            fiber
        } else {
            // SAFETY: the `make_fiber_unchecked` function is unsafe because the
            // returned fiber is unconditionally `Send` as opposed to being
            // conditionally send depending on the argument (in this case
            // `self.0`). This `async` function, however, is conditionally
            // `Send` depending on `self`, in this case `StoreContextMut<T>`,
            // which is already going to be conditionally `Send` depending on
            // `T`.
            //
            // The returned fiber is possibly stored within the `Store<T>` as
            // well. If `T: Send` then that's fine and everything's dandy. If
            // `T: !Send`, however, then the store is already not-`Send` meaning
            // that putting more actually-not-`Send` things inside of it isn't
            // an issue.
            //
            // The main issue here is that the returned fiber effectively can't
            // get transferred outside the context of the store. That's an
            // implementation detail we'll have to rely on, but is currently
            // true.
            unsafe {
                fiber::make_fiber_unchecked(self.0, move |store| {
                    loop {
                        let Some(item) = store.concurrent_state_mut()?.worker_item.take() else {
                            bail_bug!("worker_item not present when resuming fiber")
                        };
                        match item {
                            WorkerItem::GuestCall(call) => handle_guest_call(store, call)?,
                            WorkerItem::Function(fun) => fun.into_inner()(store)?,
                        }

                        store.suspend(SuspendReason::NeedWork)?;
                    }
                })?
            }
        };

        // No fallible operation between taking the fiber and guarding it in
        // resume_fiber; the current thread was forced before the take.
        self.0
            .concurrent_state_mut_already_forced_current_thread()
            .worker_item = Some(item);

        self.0.resume_fiber(worker).await
    }

    /// Wrap the specified host function in a future which will call it, passing
    /// it an `&Accessor<T>`.
    ///
    /// See the `Accessor` documentation for details.
    pub(crate) fn wrap_call<F, R>(self, closure: F) -> impl Future<Output = Result<R>> + 'static
    where
        T: 'static,
        F: FnOnce(&Accessor<T>) -> Pin<Box<dyn Future<Output = Result<R>> + Send + '_>>
            + Send
            + Sync
            + 'static,
        R: Send + Sync + 'static,
    {
        let token = StoreToken::new(self);
        async move {
            let mut accessor = Accessor::new(token);
            closure(&mut accessor).await
        }
    }

    pub(crate) async fn start_instance(
        &mut self,
        instance: ModuleInstance,
        callee: Option<RuntimeInstance>,
    ) -> Result<ModuleInstance> {
        let (tx, rx) = oneshot::channel();
        let token = StoreToken::new(self.as_context_mut());
        self.0.queue_task(move |store| {
            _ = tx.send(
                super::instance::start_raw(&mut token.as_context_mut(store), instance, callee)
                    .map(|()| instance),
            );
            Ok(())
        })?;
        self.as_context_mut()
            .run_concurrent_trap_on_idle(async |_| {
                rx.await
                    .map_err(|_| format_err!("oneshot channel canceled"))
            })
            .await??
    }
}

/// Return value of [`StoreOpaque::host_task_create`].
///
/// This is an `Option` to handle the dynamic `store.concurrency_support()`
/// property. When present this records the guest thread to restore when the
/// host call exits. The corresponding [`HostTask`] will need to be lazily
/// created if needed via [`StoreOpaque::materialize_host_task_id`].
pub type EnteredHostTask = Option<QualifiedThreadId>;

impl StoreOpaque {
    /// Returns the currently-running thread, promoting any deferred lazy guest
    /// thread into a fully-materialized `CurrentThread`. Deferred [`HostTask`]s
    /// are not materialized.
    #[inline]
    pub(crate) fn current_thread(&mut self) -> Result<CurrentThread> {
        // Without concurrency support there is nothing to force.
        if !self.concurrency_support() {
            return Ok(CurrentThread::None);
        }

        // If the JIT-visible current thread isn't a deferred thread then
        // `ConcurrentState` is already up to date.
        if !self
            .vm_store_context_mut()
            .current_thread_mut()
            .is_deferred()
        {
            return Ok(self
                .concurrent_state_mut_already_forced_current_thread()
                .unforced_current_thread);
        }

        self.force_deferred_current_thread()
    }

    /// Slow path of [`Self::current_thread`]: promote the deferred lazy
    /// thread into a fully-materialized `CurrentThread`.
    #[cold]
    fn force_deferred_current_thread(&mut self) -> Result<CurrentThread> {
        // The component instance whose adapters pushed the deferred frames; all
        // frames in a guest-to-guest, sync-to-sync call chain of fused adapters
        // live within a single `wasmtime::component::Instance` (because
        // cross-`wasmtime::component::Instance` calls don't go through fused
        // adapters), and guest code only ever runs as a guest thread, so the
        // chain's base thread is already materialized in `ConcurrentState` and
        // we can get the `ComponentInstanceId` shared by the whole chain from
        // here.
        let state = self.concurrent_state_mut_without_forcing_current_thread();
        let id = match state.unforced_current_thread.guest_task() {
            Some(task) => state.get_mut(task)?.instance.instance,
            None => bail_bug!("deferred component-model thread with non-guest base"),
        };

        // Collect the deferred frames pushed inline by fused adapters, walking
        // the `parent` chain from innermost to the base.
        let mut frames = Vec::new();
        let mut cur = *self.vm_store_context_mut().current_thread_mut();
        while let Some(ptr) = cur.as_deferred() {
            // SAFETY: `ptr` points at a `VMDeferredThread` living in a fused
            // adapter's stack frame that is suspended below us on the stack
            // (mid-call, waiting for this nested call to return), so the
            // referent is still valid and exclusively ours to read.
            let deferred = unsafe { ptr.as_non_null().as_ref() };
            frames.push((
                deferred.callee_async != 0,
                deferred.callee_instance,
                deferred.saved_context,
            ));
            cur = deferred.parent;
        }

        // Mark the current thread forced *before* replaying so that any
        // reentrant `force_current_thread` call short-circuits via the
        // non-deferred path above.
        *self.vm_store_context_mut().current_thread_mut() = VMLazyThread::forced();

        // Save the current context, as we need to overwrite it while replaying
        // below.
        let current_context = *self.vm_store_context_mut().component_context_mut();

        // Replay the deferred `enter_guest_sync_call`s outermost-first so that
        // the resulting `ConcurrentState` matches what the non-deferred path
        // would have otherwise produced.
        for (callee_async, callee_instance, saved_context) in frames.into_iter().rev() {
            // Restore the caller's context slots so that we save the correct
            // values into the caller's thread, exactly as the non-deferred path
            // would have on entry.
            *self.vm_store_context_mut().component_context_mut() = saved_context;
            let callee = RuntimeInstance {
                instance: id,
                index: RuntimeComponentInstanceIndex::from_u32(callee_instance),
            };
            self.enter_guest_sync_call(callee_async, callee)?;
        }

        // Replaying done; restore the current context.
        *self.vm_store_context_mut().component_context_mut() = current_context;

        Ok(self
            .concurrent_state_mut_without_forcing_current_thread()
            .unforced_current_thread)
    }

    fn current_guest_thread(&mut self) -> Result<QualifiedThreadId> {
        match self.current_thread()?.guest() {
            Some(id) => Ok(*id),
            None => bail_bug!("current thread is not a guest thread"),
        }
    }

    // A result of `None` may indicate that this is either the top-level event
    // loop, a deferred host task, or concurrency support is disabled. In all
    // cases we don't have an ID for the task.
    pub(crate) fn current_materialized_host_task(&mut self) -> Result<Option<TableId<HostTask>>> {
        match self.current_thread()? {
            CurrentThread::Host(id) => Ok(Some(id)),
            CurrentThread::DeferredHost(_) | CurrentThread::None => Ok(None),
            _ => bail_bug!("current thread is not a host thread"),
        }
    }

    /// Returns the current host task ID, materializing a deferred host task if
    /// one is active. `None` represents a call from the top-level host.
    fn materialize_host_task_id(&mut self) -> Result<Option<TableId<HostTask>>> {
        Ok(self
            .concurrent_state_mut()?
            .materialize_current_host_task_id()?)
    }

    fn enter_sync_call(&mut self, callee: RuntimeInstance) -> Result<()> {
        log::trace!("enter sync-typed call {callee:?}");
        let state = self.instance_state(callee).concurrent_state();
        let old_do_not_suspend = state.do_not_suspend;
        state.do_not_suspend = true;

        let thread = self.current_guest_thread()?;
        let thread = self.concurrent_state_mut()?.get_mut(thread.thread)?;
        if thread.old_do_not_suspend.is_some() {
            bail_bug!("current thread already has `old_do_not_suspend` value");
        }

        thread.old_do_not_suspend = Some(old_do_not_suspend);

        Ok(())
    }

    fn exit_sync_call(&mut self, callee: RuntimeInstance) -> Result<()> {
        log::trace!("exit sync-typed call {callee:?}");
        let thread = self.current_guest_thread()?;
        let thread = self.concurrent_state_mut()?.get_mut(thread.thread)?;
        let Some(old_do_not_suspend) = thread.old_do_not_suspend.take() else {
            bail_bug!("current thread missing `old_do_not_suspend` value");
        };
        let state = self.instance_state(callee).concurrent_state();
        state.do_not_suspend = old_do_not_suspend;
        Ok(())
    }

    /// Push a `GuestTask` onto the task stack for either a sync-to-sync,
    /// guest-to-guest call or a sync host-to-guest call.
    ///
    /// This task will only be used for the purpose of handling calls to
    /// intrinsic functions; both parameter lowering and result lifting are
    /// assumed to be taken care of elsewhere.
    ///
    /// NB: for sync-to-sync, guest-to-guest calls we delay task construction in
    /// fused adapters, see `StoreOpaque::current_thread`, `VMDeferredThread`,
    /// and `lower_fact_enter_sync_call`. Make sure all this stuff stays in
    /// sync!
    pub(crate) fn enter_guest_sync_call(
        &mut self,
        callee_async_typed: bool,
        callee: RuntimeInstance,
    ) -> Result<()> {
        log::trace!("enter sync-lifted call {callee:?}");
        if !self.concurrency_support() {
            return self.enter_call_not_concurrent();
        }

        // Stash the next switch item (if any) for the duration of this call
        // since it's for the caller's caller, not for us.  We'll restore this
        // in `exit_guest_sync_call`.
        let state = self.concurrent_state_mut()?;
        let item = state.next_switch_item.take();
        state.saved_next_switch_items.push(item);

        let thread = self.current_thread()?;
        let caller = if let Some(thread) = thread.guest() {
            Caller::Guest { thread: *thread }
        } else {
            Caller::Host {
                tx: None,
                caller: self.materialize_host_task_id()?,
            }
        };

        let state = self.concurrent_state_mut()?;
        let guest_thread = GuestTask::new(
            state,
            Box::new(move |_, _| bail_bug!("cannot lower params in sync call")),
            LiftResult {
                lift: Box::new(move |_, _| bail_bug!("cannot lift result in sync call")),
                ty: TypeTupleIndex::reserved_value(),
                memory: None,
                string_encoding: StringEncoding::Utf8,
            },
            caller,
            None,
            callee,
            callee_async_typed,
            false,
        )?;

        Instance::from_wasmtime(self, callee.instance).add_guest_thread_to_instance_table(
            guest_thread.thread,
            self,
            callee.index,
        )?;
        self.set_thread(guest_thread)?;

        if !callee_async_typed {
            self.enter_sync_call(callee)?;
        }

        Ok(())
    }

    /// Pop a `GuestTask` previously pushed using `enter_guest_sync_call`.
    ///
    /// NB: for sync-to-sync, guest-to-guest calls we delay task construction in
    /// fused adapters and then when the call returns we check to see if the
    /// task's contruction was forced and if not avoid calling out of the JIT
    /// code to this function. See `lower_fact_exit_sync_call`. Make sure all
    /// this stuff stays in sync!
    pub(crate) fn exit_guest_sync_call(&mut self) -> Result<()> {
        if !self.concurrency_support() {
            return Ok(self.exit_call_not_concurrent());
        }

        let thread = match self.current_thread()?.guest() {
            Some(t) => *t,
            None => bail_bug!("expected task when exiting"),
        };
        let task = self.concurrent_state_mut()?.get_mut(thread.task)?;
        let instance = task.instance;

        let caller = match &task.caller {
            &Caller::Guest { thread } => thread.into(),
            &Caller::Host { caller, .. } => caller
                .map(CurrentThread::Host)
                .unwrap_or(CurrentThread::None),
        };
        task.lift_result = None;
        task.exited = true;
        let async_typed = task.async_typed;

        if !async_typed {
            self.exit_sync_call(instance)?;
        }

        self.set_thread(caller)?;

        log::trace!("exit sync-lifted call {instance:?}");

        if async_typed {
            // If we're async-typed, returning control to our caller won't help
            // resolve any outstanding sync-typed call which might be in
            // progress, so we may need to switch or trap before exiting this
            // thread:
            self.switch_or_trap_if_may_not_suspend(instance)?;
        }

        self.cleanup_thread(thread, instance)?;

        self.concurrent_state_mut()?
            .decrement_task_ref_count(thread.task)?;

        let state = self.concurrent_state_mut()?;
        let Some(item) = state.saved_next_switch_items.pop() else {
            bail_bug!("unable to pop from `saved_next_switch_items`");
        };
        if let Some(item) = mem::replace(&mut state.next_switch_item, item) {
            // Stash it back in the store to ensure it's cleaned up on store
            // drop:
            state.push_high_priority(item);
            bail_bug!("`next_switch_item` unexpectedly already set");
        }

        Ok(())
    }

    /// Similar to `enter_guest_sync_call` except for when the guest makes a
    /// transition to the host.
    ///
    /// This initially records a deferred host call. A full [`HostTask`] should
    /// be allocated later if needed via
    /// [`StoreOpaque::materialize_host_task_id`].
    pub(crate) fn host_task_create(&mut self) -> Result<EnteredHostTask> {
        if !self.concurrency_support() {
            self.enter_call_not_concurrent()?;
            return Ok(None);
        }
        let caller = self.current_guest_thread()?;
        log::trace!("new deferred host task with caller {caller:?}");

        self.set_thread(CurrentThread::DeferredHost(caller))?;
        let state = self.concurrent_state_mut()?;
        debug_assert!(state.deferred_host_call_context.is_none());
        state.deferred_host_call_context = Some(CallContext::default());
        state.debug_assert_deferred_host_invariant();
        Ok(Some(caller))
    }

    /// Dual of `host_task_create` and signifies that the host has finished and
    /// will be cleaned up.
    ///
    /// Note that this isn't invoked when the host is invoked asynchronously and
    /// the host isn't complete yet. In that situation the host task persists
    /// and will be cleaned up separately in `subtask_drop`
    pub(crate) fn host_task_delete(
        &mut self,
        original_task: EnteredHostTask,
        materialized_task: Option<TableId<HostTask>>,
    ) -> Result<()> {
        match original_task {
            Some(caller) => {
                self.set_thread(caller)?;
                if materialized_task.is_none() {
                    let state = self.concurrent_state_mut()?;
                    let context = state
                        .deferred_host_call_context
                        .take()
                        .expect("deferred host call context should be present");
                    debug_assert!(context.is_empty());
                    state.debug_assert_deferred_host_invariant();
                }
                log::trace!(
                    "delete host task with caller {original_task:?} and materialized as {materialized_task:?}"
                );
                if let Some(task) = materialized_task {
                    self.concurrent_state_mut()?.delete_host_task(task)?;
                }
            }
            None => {
                debug_assert!(materialized_task.is_none());
                self.exit_call_not_concurrent();
            }
        }
        Ok(())
    }

    /// Helper function to retrieve the `InstanceState` for the
    /// specified instance.
    fn instance_state(&mut self, instance: RuntimeInstance) -> &mut InstanceState {
        self.component_instance_mut(instance.instance)
            .instance_state(instance.index)
    }

    /// Configure the currently running `thread`.
    ///
    /// This will save off any state necessary for the previous thread, if
    /// applicable, and then it'll additionally update state for `thread` if
    /// needed too.
    pub(crate) fn set_thread(&mut self, thread: impl Into<CurrentThread>) -> Result<CurrentThread> {
        let thread = thread.into();
        let state = self.concurrent_state_mut()?;
        state.debug_assert_deferred_host_invariant();
        let old_thread = mem::replace(&mut state.unforced_current_thread, thread);

        state.handle_thread_switch(old_thread, thread)?;

        // First thing to do after swapping threads is updating the context
        // slots for this thread within the store. This restores the behavior of
        // `context.{get,set}`. This involves taking the old state out of the
        // store, saving it in the thread that's being swapped from, and doing
        // the inverse for the new thread. When debug assertions are enabled
        // this also leaves behind sentinel values to try to uncover bugs where
        // this may be forgotten.
        if let Some(old_thread) = old_thread.guest() {
            let old_context = *self.vm_store_context_mut().component_context_mut();
            self.concurrent_state_mut()?
                .get_mut(old_thread.thread)?
                .context = old_context;
        }
        if cfg!(debug_assertions) {
            *self.vm_store_context_mut().component_context_mut() =
                [u32::MAX; NUM_COMPONENT_CONTEXT_SLOTS];
        }
        if let Some(thread) = thread.guest() {
            let thread = self.concurrent_state_mut()?.get_mut(thread.thread)?;
            let context = thread.context;
            if cfg!(debug_assertions) {
                thread.context = [u32::MAX; NUM_COMPONENT_CONTEXT_SLOTS];
            }
            *self.vm_store_context_mut().component_context_mut() = context;
        }

        // Keep the JIT-visible current-thread pointer in sync.
        *self.vm_store_context_mut().current_thread_mut() = if thread.is_none() {
            VMLazyThread::none()
        } else {
            VMLazyThread::forced()
        };

        Ok(old_thread)
    }

    /// Call `switch_if_may_not_suspend` and trap if it returns `false`.
    fn switch_or_trap_if_may_not_suspend(&mut self, instance: RuntimeInstance) -> Result<()> {
        if self.switch_if_may_not_suspend(instance)? {
            Ok(())
        } else {
            Err(Trap::CannotBlockSyncTask.into())
        }
    }

    /// Check if the specified instance has a sync-typed call in progress; if so
    /// attempt to switch to another ready thread for that instance, and if no
    /// such thread exists, return false.
    fn switch_if_may_not_suspend(&mut self, instance: RuntimeInstance) -> Result<bool> {
        // Call this for the side effect of forcing any deferred task creation,
        // which may influence the value of `ConcurrentState::do_not_suspend`
        // below:
        self.concurrent_state_mut()?;

        Ok(!self.concurrency_support()
            || !self
                .instance_state(instance)
                .concurrent_state()
                .do_not_suspend
            || self
                .concurrent_state_mut()?
                .promote_instance_local_thread_work_item(instance)?)
    }

    /// Record that we're about to enter a (sub-)component instance which does
    /// not support more than one concurrent, stackful activation, meaning it
    /// cannot be entered again until the next call returns.
    fn enter_instance(&mut self, instance: RuntimeInstance) {
        log::trace!("enter {instance:?}");
        self.instance_state(instance)
            .concurrent_state()
            .do_not_enter = true;
    }

    /// Record that we've exited a (sub-)component instance previously entered
    /// with `Self::enter_instance` and then calls `Self::partition_pending`.
    /// See the documentation for the latter for details.
    fn exit_instance(&mut self, instance: RuntimeInstance) -> Result<()> {
        log::trace!("exit {instance:?}");
        self.instance_state(instance)
            .concurrent_state()
            .do_not_enter = false;
        self.partition_pending(instance)
    }

    /// Iterate over `InstanceState::pending`, moving any ready items into the
    /// "high priority" work item queue.
    ///
    /// Also, notify `ConcurrentState::ready_for_concurrent_call_waker` if
    /// present.
    ///
    /// See `GuestCall::is_ready` for details.
    fn partition_pending(&mut self, instance: RuntimeInstance) -> Result<()> {
        for (thread, kind) in
            mem::take(&mut self.instance_state(instance).concurrent_state().pending).into_iter()
        {
            let call = GuestCall { thread, kind };
            if call.is_ready(self)? {
                self.concurrent_state_mut()?
                    .push_high_priority(WorkItem::GuestCall { instance, call });
            } else {
                self.instance_state(instance)
                    .concurrent_state()
                    .pending
                    .insert(call.thread, call.kind);
            }
        }

        if let Some(waker) = self
            .concurrent_state_mut()?
            .ready_for_concurrent_call_waker
            .take()
        {
            waker.wake();
        }

        Ok(())
    }

    /// Implements the `backpressure.{inc,dec}` intrinsics.
    pub(crate) fn backpressure_modify(
        &mut self,
        caller_instance: RuntimeInstance,
        modify: impl FnOnce(u16) -> Option<u16>,
    ) -> Result<()> {
        let state = self.instance_state(caller_instance).concurrent_state();
        let old = state.backpressure;
        let new = modify(old).ok_or_else(|| Trap::BackpressureOverflow)?;
        state.backpressure = new;

        if old > 0 && new == 0 {
            // Backpressure was previously enabled and is now disabled; move any
            // newly-eligible guest calls to the "high priority" queue.
            self.partition_pending(caller_instance)?;
        }

        Ok(())
    }

    /// Resume the specified fiber, giving it exclusive access to the specified
    /// store.
    fn resume_fiber(
        &mut self,
        fiber: StoreFiber<'static>,
    ) -> impl Future<Output = Result<()>> + '_ {
        let mut resuming = ResumingFiber {
            store: self,
            fiber: Some(fiber),
        };
        async move {
            let old_thread = resuming.store.current_thread()?;
            log::trace!("resume_fiber: save current thread {old_thread:?}");

            let fiber = resuming.fiber.take().unwrap();
            resuming.fiber = fiber::resolve_or_release(resuming.store, fiber).await?;

            resuming.store.set_thread(old_thread)?;

            let state = resuming.store.concurrent_state_mut()?;

            if let Some(ot) = old_thread.guest() {
                state.set_thread_running(ot.thread)?;
            }
            log::trace!("resume_fiber: restore current thread {old_thread:?}");

            if resuming.fiber.is_some() {
                log::trace!("resume_fiber: suspend reason {:?}", &state.suspend_reason);
                // See the `SuspendReason` documentation for what each case means.
                let reason = match state.suspend_reason.take() {
                    Some(r) => r,
                    None => bail_bug!("suspend reason missing when resuming fiber"),
                };
                match reason {
                    SuspendReason::NeedWork => {
                        if state.worker.is_none() {
                            state.worker = resuming.fiber.take();
                        }
                    }
                    SuspendReason::Yielding { thread } => {
                        let slot = &mut state.get_mut(thread.thread)?.state;
                        slot.check_no_fiber()?;
                        *slot = GuestThreadState::Ready {
                            fiber: resuming.fiber.take().unwrap(),
                        };
                        let instance = state.get_mut(thread.task)?.instance;
                        state.push_low_priority(WorkItem::ResumeThread { instance, thread });
                    }
                    SuspendReason::ExplicitlySuspending { thread } => {
                        let slot = &mut state.get_mut(thread.thread)?.state;
                        slot.check_no_fiber()?;
                        *slot = GuestThreadState::Suspended(resuming.fiber.take().unwrap());
                    }
                    SuspendReason::Waiting { set, thread } => {
                        match state.get_mut(set)?.waiting.entry(thread) {
                            Entry::Vacant(entry) => {
                                entry.insert(WaitMode::Fiber(resuming.fiber.take().unwrap()));
                            }
                            Entry::Occupied(_) => {
                                bail_bug!("entry unexpectedly already exists for {thread:?}");
                            }
                        }
                    }
                    SuspendReason::YieldingToSubtask { thread } => {
                        // In this case, the thread has either invoked or sent a
                        // cancel request to a subtask, and is now yielding to that
                        // subtask.  According to the CM spec, that subtask may only
                        // yield back to the original thread the first time it
                        // suspends or exits (or a thread that it has resumed
                        // suspends or exits, etc.), which we ensure by setting
                        // `ConcurrentState::next_switch_item` here.

                        let instance = state.get_mut(thread.task)?.instance;
                        if state.next_switch_item.is_some() {
                            // This should be unreachable per the save/restore code
                            // in `Self::suspend`.
                            bail_bug!(
                                "`ConcurrentState::next_switch_item` was already `Some(_)` when \
                             a thread wanted to wait on a subtask"
                            );
                        }
                        let item = state.next_switch_item.insert(WorkItem::ResumeFiber {
                            instance,
                            thread,
                            fiber: resuming.fiber.take().unwrap(),
                        });
                        log::trace!("set next switch item to {item:?}");
                    }
                };
            } else {
                log::trace!("resume_fiber: fiber has exited");
            }

            Ok(())
        }
    }

    /// Suspend the current fiber, storing the reason in
    /// `ConcurrentState::suspend_reason` to indicate the conditions under which
    /// it should be resumed.
    ///
    /// See the `SuspendReason` documentation for details.
    fn suspend(&mut self, reason: SuspendReason) -> Result<()> {
        log::trace!("suspend fiber: {reason:?}");

        let state = self.concurrent_state_mut()?;

        // If we're yielding or waiting on behalf of a guest thread, we'll be
        // overwriting the current thread, so save it now and restore it once
        // we've resumed.
        //
        // Also, if we're yielding to a subtask, we're about to overwrite
        // `ConcurrentState::next_switch_item`, so also save and restore that.
        let (save_and_restore_thread, save_and_restore_next_switch_item) = match &reason {
            SuspendReason::Yielding { .. }
            | SuspendReason::Waiting { .. }
            | SuspendReason::ExplicitlySuspending { .. } => {
                // If there's a thread waiting for this subtask to suspend, this
                // is a good time to switch back to it.
                if state.switch_item.is_none() {
                    state.take_next_switch_item()?;
                }

                (true, false)
            }
            SuspendReason::YieldingToSubtask { .. } => (true, true),
            SuspendReason::NeedWork => (false, false),
        };

        let old_next_switch_item = if save_and_restore_next_switch_item {
            Some(state.save_next_switch_item()?)
        } else {
            None
        };

        let old_guest_thread = if save_and_restore_thread {
            self.current_thread()?
        } else {
            CurrentThread::None
        };

        let waiting_set = match &reason {
            SuspendReason::Waiting { set, .. } => Some(*set),
            _ => None,
        };

        let suspend_reason = &mut self.concurrent_state_mut()?.suspend_reason;
        assert!(suspend_reason.is_none());
        *suspend_reason = Some(reason);

        // We'll panic if we call `Self::with_blocking` when the fiber is being
        // disposed, so check for that and exit ASAP if appropriate.
        if !self.fiber_async_state_mut().can_block() {
            return Err(format_err!("future dropped"));
        }

        // Count this thread as waiting on `waiting_set`, if any, for as long as
        // it's suspended.
        if let Some(set) = waiting_set {
            self.concurrent_state_mut()?.get_mut(set)?.num_waiting += 1;
        }

        self.with_blocking(|_, cx| cx.suspend(StoreFiberYield::ReleaseStore))?;

        if let Some(set) = waiting_set {
            self.concurrent_state_mut()?.get_mut(set)?.stop_waiting()?;
        }

        if save_and_restore_thread {
            self.set_thread(old_guest_thread)?;
        }

        if let Some(item) = old_next_switch_item {
            let state = self.concurrent_state_mut()?;
            state.restore_next_switch_item(item)?;
        }

        Ok(())
    }

    fn wait_for_event(
        &mut self,
        caller_instance: RuntimeInstance,
        waitable: Waitable,
    ) -> Result<()> {
        let caller = self.current_guest_thread()?;
        let state = self.concurrent_state_mut()?;

        waitable.trap_if_in_waitable_set(state)?;

        let set = state.get_mut(caller.thread)?.sync_call_set;
        waitable.join(state, Some(set))?;

        self.switch_or_trap_if_may_not_suspend(caller_instance)?;

        self.suspend(SuspendReason::Waiting {
            set,
            thread: caller,
        })?;
        let state = self.concurrent_state_mut()?;

        waitable.join(state, None)
    }

    /// Cleans up the data structures backing the `guest_thread` specified,
    /// removing it from the internal tables of `runtime_instance` as well.
    ///
    /// This function is used whenever a guest thread has fully exited and
    /// completed. This'll clean up the associated `GuestThread` structure and
    /// related resources it contains.
    ///
    /// Other functionality that this implements is:
    ///
    /// * This will perform conditional cleanup of the `GuestTask` that owns
    ///   this thread if `cleanup_task` is `CleanupTask::Yes`.
    /// * If there are no more threads in the `GuestTask` that this thread is
    ///   associated with, and if the task hasn't produced a result (e.g. it's not
    ///   returned or cancelled), then a trap will be raised that a result
    ///   wasn't ever produced.
    /// * If this task is finished, meaning the top-level thread exited and
    ///   additionally it's been returned or cancelled, then this will handle
    ///   management of the store's "active interesting tasks" counter.
    ///
    /// Effectively this is intended to be a "narrow waist" through which many
    /// destruction operations are funneled through.
    fn cleanup_thread(
        &mut self,
        guest_thread: QualifiedThreadId,
        runtime_instance: RuntimeInstance,
    ) -> Result<()> {
        let state = self.concurrent_state_mut()?;
        // If we never suspended, we never had a chance to deliver a subtask
        // status update, if any, to our caller, so we do that here:
        state.take_next_switch_item()?;
        let thread_data = state.get_mut(guest_thread.thread)?;
        let sync_call_set = thread_data.sync_call_set;
        if let Some(guest_id) = thread_data.instance_rep {
            self.instance_state(runtime_instance)
                .thread_handle_table()
                .guest_thread_remove(guest_id)?;
        }
        let state = self.concurrent_state_mut()?;

        // Clean up any pending subtasks in the sync_call_set
        for waitable in mem::take(&mut state.get_mut(sync_call_set)?.ready) {
            if let Some(Event::Subtask {
                status: Status::Returned | Status::ReturnCancelled,
            }) = waitable.common(state)?.event
            {
                waitable.delete_from(state)?;
            }
        }

        state.delete(guest_thread.thread)?;
        state.delete(sync_call_set)?;
        let task = state.get_mut(guest_thread.task)?;
        task.threads.remove(&guest_thread.thread);

        if task.threads.is_empty() && !task.returned_or_cancelled() {
            bail!(Trap::NoAsyncResult);
        }

        if !task.decremented_interesting_task_count && task.exited && task.returned_or_cancelled() {
            task.decremented_interesting_task_count = true;

            debug_assert!(state.interesting_tasks > 0);
            state.interesting_tasks -= 1;
            if state.interesting_tasks == 0
                && let Some(waker) = state.interesting_tasks_empty_waker.take()
            {
                waker.wake();
            }
        }

        state.decrement_task_ref_count(guest_thread.task)?;

        Ok(())
    }

    /// Performs cancellation of the `guest_task` specified with the
    /// precondition that the task hasn't lowered its parameters.
    ///
    /// In this situation the task hasn't ever been started meaning it hasn't
    /// actually run any wasm code yet. This requires cleaning up metadata such
    /// as thread information attached to the task.
    ///
    /// The main two entrypoints for this function are:
    ///
    /// * Task cancellation via `subtask.cancel`, the intrinsic.
    /// * Dropping a host `call_async` future which needs to cancel the task
    ///   because it cannot reference its parameters any more.
    fn cancel_guest_subtask_without_lowered_parameters(
        &mut self,
        caller_instance: RuntimeInstance,
        guest_task: TableId<GuestTask>,
    ) -> Result<()> {
        let concurrent_state = self.concurrent_state_mut()?;
        let task = concurrent_state.get_mut(guest_task)?;
        assert!(!task.already_lowered_parameters());
        // The task is in a `starting` state, meaning it hasn't run at
        // all yet.  Here we update its fields to indicate that it is
        // ready to delete immediately once `subtask.drop` is called.
        task.lower_params = None;
        task.lift_result = None;
        task.exited = true;
        let instance = task.instance;

        // Clean up the thread within this task as it's now never going
        // to run.
        assert_eq!(1, task.threads.len());
        let thread = *task.threads.iter().next().unwrap();
        self.cleanup_thread(
            QualifiedThreadId {
                task: guest_task,
                thread,
            },
            caller_instance,
        )?;

        // Not yet started; cancel and remove from pending
        let pending = &mut self.instance_state(instance).concurrent_state().pending;
        let pending_count = pending.len();
        pending.retain(|thread, _| thread.task != guest_task);
        // If there were no pending threads for this task, we're in an error state
        if pending.len() == pending_count {
            bail!(Trap::SubtaskCancelAfterTerminal);
        }
        Ok(())
    }

    /// Used by `ResourceTables` to record the scope of a borrow to get undone
    /// in the future.
    pub(crate) fn current_scope(&mut self) -> Result<Option<CurrentScope>> {
        if !self.concurrency_support() {
            return Ok(self
                .current_scope_id_not_concurrent()?
                .map(|id| CurrentScope::Id(Scope::Id(id))));
        }

        Ok(match self.current_thread()? {
            CurrentThread::Guest(id) => Some(CurrentScope::Id(Scope::Id(id.task.rep()))),
            CurrentThread::Host(id) => Some(CurrentScope::Id(Scope::HostId(id.rep()))),
            CurrentThread::DeferredHost(_) => Some(CurrentScope::DeferredHost),
            CurrentThread::None => return Ok(None),
        })
    }

    pub(crate) fn queue_task(
        &mut self,
        task: impl FnOnce(&mut dyn VMStore) -> Result<()> + Send + 'static,
    ) -> Result<()> {
        self.concurrent_state_mut()?
            .push_high_priority(WorkItem::WorkerFunction(AlwaysMut::new(Box::new(task))));
        Ok(())
    }

    /// Used in `poll_until` just prior to trapping due to a "deadlock"
    /// condition.
    ///
    /// This helps us distinguish between a simple deadlock condition (where no
    /// work is available for the event loop to do, nor is there any way for new
    /// work to be added) and a "cannot block sync task" condition where at
    /// least once instance has an outstanding sync-typed task running, in which
    /// case we'll trap with a different error message.
    fn any_may_not_suspend(&mut self) -> Result<bool> {
        // Note that this currently requires a linear search across the whole
        // `ConcurrentState::table`.  We _could_ optimize that, but since (1)
        // this function is only used when trapping, (2) the only thing you can
        // really do with a store that's been poisoned by a trap is drop it, and
        // (3) we must do a linear search through the table when dropping a
        // store anyway to dispose of fibers, it's reasonable for us to also do
        // a linear search here.
        Ok(self
            .concurrent_state_mut()?
            .table
            .get_mut()
            .iter_mut()
            .filter_map(|(_, entry)| {
                if let Some(task) = entry.downcast_ref::<GuestTask>() {
                    Some(task.instance)
                } else {
                    None
                }
            })
            .collect::<Vec<_>>()
            .into_iter()
            .any(|instance| {
                self.instance_state(instance)
                    .concurrent_state()
                    .do_not_suspend
            }))
    }
}

impl Instance {
    /// Get the next pending event for the specified task and (optional)
    /// waitable set, along with the waitable handle if applicable.
    fn get_event(
        self,
        store: &mut StoreOpaque,
        guest_task: TableId<GuestTask>,
        set: Option<TableId<WaitableSet>>,
        cancellable: bool,
    ) -> Result<Option<(Event, Option<(Waitable, u32)>)>> {
        let state = store.concurrent_state_mut()?;

        let task = state.get_mut(guest_task)?;
        let event = &mut task.event;
        if let Some(ev) = event
            && (cancellable || !matches!(ev, Event::Cancelled))
        {
            log::trace!("deliver event {ev:?} to {guest_task:?}");

            if matches!(ev, Event::Cancelled) {
                task.cancel_request_delivered = true;
            }

            let ev = *ev;
            *event = None;
            return Ok(Some((ev, None)));
        }

        let set = match set {
            Some(set) => set,
            None => return Ok(None),
        };
        let waitable = match state.get_mut(set)?.ready.pop_first() {
            Some(v) => v,
            None => return Ok(None),
        };

        let common = waitable.common(state)?;
        let handle = match common.handle {
            Some(h) => h,
            None => bail_bug!("handle not set when delivering event"),
        };
        let event = match common.event.take() {
            Some(e) => e,
            None => bail_bug!("event not set when delivering event"),
        };

        log::trace!(
            "deliver event {event:?} to {guest_task:?} for {waitable:?} (handle {handle}); set {set:?}"
        );

        waitable.on_delivery(store, self, event)?;

        Ok(Some((event, Some((waitable, handle)))))
    }

    /// Handle the `CallbackCode` returned from an async-lifted export or its
    /// callback.
    ///
    /// If this returns `Ok(Some(call))`, then `call` should be run immediately
    /// using `handle_guest_call`.
    fn handle_callback_code(
        self,
        store: &mut StoreOpaque,
        guest_thread: QualifiedThreadId,
        runtime_instance: RuntimeComponentInstanceIndex,
        code: u32,
    ) -> Result<()> {
        let (code, set) = unpack_callback_code(code);

        log::trace!("received callback code from {guest_thread:?}: {code} (set: {set})");

        let state = store.concurrent_state_mut()?;

        state.take_next_switch_item()?;

        let get_set = |store: &mut StoreOpaque, handle| -> Result<_> {
            let set = store
                .instance_state(self.runtime_instance(runtime_instance))
                .handle_table()
                .waitable_set_rep(handle)?;

            Ok(TableId::<WaitableSet>::new(set))
        };

        match code {
            callback_code::EXIT => {
                log::trace!("implicit thread {guest_thread:?} completed");
                let task = store.concurrent_state_mut()?.get_mut(guest_thread.task)?;
                task.exited = true;
                task.callback = None;

                let runtime_instance = self.runtime_instance(runtime_instance);

                // Since we're async-typed, returning control to our caller
                // won't help resolve any outstanding sync-typed call which
                // might be in progress, so we may need to switch or trap before
                // exiting this thread:
                store.switch_or_trap_if_may_not_suspend(runtime_instance)?;

                store.cleanup_thread(guest_thread, runtime_instance)?;
            }
            callback_code::YIELD => {
                // Set `GuestTask::wake_on_cancel` to allow `subtask.cancel` to
                // promote this thread if appropriate.
                let old = state
                    .get_mut(guest_thread.thread)?
                    .wake_on_cancel
                    .replace(WakeOnCancel::Yielding);
                if !old.is_none() {
                    bail_bug!("thread unexpectedly had wake_on_cancel set");
                }

                // If an `Event::Cancelled` is pending in `GuestTask::event`
                // when this call is delivered, we'll deliver that; otherwise,
                // we'll deliver `Event::None`.  Note that the `Event::None` is
                // deliberately not stored in `GuestTask::event` since that's
                // shared by all threads of the task and the `Event::None` is
                // only for this thread.
                let call = GuestCall {
                    thread: guest_thread,
                    kind: GuestCallKind::DeliverEvent {
                        instance: self,
                        set: None,
                    },
                };
                // Push this thread onto the "low priority" queue so it runs
                // after any other threads have had a chance to run.
                state.push_low_priority(WorkItem::GuestCall {
                    instance: self.runtime_instance(runtime_instance),
                    call,
                });
            }
            callback_code::WAIT => {
                let set = get_set(store, set)?;
                self.wait_with_callback(store, guest_thread, set)?;
            }
            _ => bail!(Trap::UnsupportedCallbackCode),
        }

        Ok(())
    }

    /// Start waiting on `set` with `guest_thread`, which uses a callback.
    ///
    /// The thread will receive an event from `set` via a
    /// `GuestCallKind::DeliverEvent` work item, either immediately if one is
    /// already available, or once one is published.
    fn wait_with_callback(
        self,
        store: &mut StoreOpaque,
        guest_thread: QualifiedThreadId,
        set: TableId<WaitableSet>,
    ) -> Result<()> {
        let state = store.concurrent_state_mut()?;
        // This thread's wait ends when that work item is handled in
        // `handle_guest_call`.
        state.get_mut(set)?.num_waiting += 1;

        if state.get_mut(guest_thread.task)?.event.is_some()
            || !state.get_mut(set)?.ready.is_empty()
        {
            // An event is immediately available; deliver it ASAP.
            let instance = state.get_mut(guest_thread.task)?.instance;
            state.push_high_priority(WorkItem::GuestCall {
                instance,
                call: GuestCall {
                    thread: guest_thread,
                    kind: GuestCallKind::DeliverEvent {
                        instance: self,
                        set: Some(set),
                    },
                },
            });
            return Ok(());
        }

        // No event is immediately available, so this thread is about to block.
        // Like any other suspension point (e.g. the callback `EXIT` path), if
        // a sync-typed call is in progress for this instance we must first
        // switch to another ready thread of that instance, or trap if there is
        // none.
        let instance = store
            .concurrent_state_mut()?
            .get_mut(guest_thread.task)?
            .instance;
        store.switch_or_trap_if_may_not_suspend(instance)?;

        // Register to be woken up when an event is published for this waitable
        // set.
        //
        // Here we also set `GuestTask::wake_on_cancel` which allows
        // `subtask.cancel` to interrupt the wait.
        let state = store.concurrent_state_mut()?;
        let old = state
            .get_mut(guest_thread.thread)?
            .wake_on_cancel
            .replace(WakeOnCancel::Waiting(set));
        if !old.is_none() {
            bail_bug!("thread unexpectedly had wake_on_cancel set");
        }
        let old = state
            .get_mut(set)?
            .waiting
            .insert(guest_thread, WaitMode::Callback(self));
        if !old.is_none() {
            bail_bug!("set's waiting set already had this thread registered");
        }
        Ok(())
    }

    /// Stage the specified guest call as the work item to run next, to be
    /// started as soon as backpressure and/or reentrance rules allow.
    ///
    /// SAFETY: The raw pointer arguments must be valid references to guest
    /// functions (with the appropriate signatures) when the closures staged by
    /// this function are called.
    unsafe fn stage_call<T: 'static>(
        self,
        mut store: StoreContextMut<T>,
        guest_thread: QualifiedThreadId,
        callee: SendSyncPtr<VMFuncRef>,
        param_count: usize,
        result_count: usize,
        async_: bool,
        callback: Option<SendSyncPtr<VMFuncRef>>,
        post_return: Option<SendSyncPtr<VMFuncRef>>,
        host_caller: bool,
    ) -> Result<()> {
        /// Return a closure which will call the specified function in the scope
        /// of the specified task.
        ///
        /// This will use `GuestTask::lower_params` to lower the parameters, but
        /// will not lift the result; instead, it returns a
        /// `[MaybeUninit<ValRaw>; MAX_FLAT_PARAMS]` from which the result, if
        /// any, may be lifted.  Note that an async-lifted export will have
        /// returned its result using the `task.return` intrinsic (or not
        /// returned a result at all, in the case of `task.cancel`), in which
        /// case the "result" of this call will either be a callback code or
        /// nothing.
        ///
        /// SAFETY: `callee` must be a valid `*mut VMFuncRef` at the time when
        /// the returned closure is called.
        unsafe fn make_call<T: 'static>(
            store: StoreContextMut<T>,
            guest_thread: QualifiedThreadId,
            callee: SendSyncPtr<VMFuncRef>,
            param_count: usize,
            result_count: usize,
        ) -> impl FnOnce(&mut dyn VMStore) -> Result<[MaybeUninit<ValRaw>; MAX_FLAT_PARAMS]>
        + Send
        + Sync
        + 'static
        + use<T> {
            let token = StoreToken::new(store);
            move |store: &mut dyn VMStore| {
                let mut storage = [MaybeUninit::uninit(); MAX_FLAT_PARAMS];

                store
                    .concurrent_state_mut()?
                    .set_thread_running(guest_thread.thread)?;
                let task = store.concurrent_state_mut()?.get_mut(guest_thread.task)?;
                let lower = match task.lower_params.take() {
                    Some(l) => l,
                    None => bail_bug!("lower_params missing"),
                };

                lower(store, &mut storage[..param_count])?;

                let mut store = token.as_context_mut(store);

                // SAFETY: Per the contract documented in `make_call's`
                // documentation, `callee` must be a valid pointer.
                unsafe {
                    crate::Func::call_unchecked_raw(
                        &mut store,
                        callee.as_non_null(),
                        NonNull::new(
                            &mut storage[..param_count.max(result_count)]
                                as *mut [MaybeUninit<ValRaw>] as _,
                        )
                        .unwrap(),
                        UncaughtException::Trap,
                    )?;
                }

                Ok(storage)
            }
        }

        // SAFETY: Per the contract described in this function documentation,
        // the `callee` pointer which `call` closes over must be valid when
        // called by the closure we queue below.
        let call = unsafe {
            make_call(
                store.as_context_mut(),
                guest_thread,
                callee,
                param_count,
                result_count,
            )
        };

        let callee_instance = store
            .0
            .concurrent_state_mut()?
            .get_mut(guest_thread.task)?
            .instance;

        let fun = if callback.is_some() {
            assert!(async_);

            Box::new(move |store: &mut dyn VMStore| {
                self.add_guest_thread_to_instance_table(
                    guest_thread.thread,
                    store,
                    callee_instance.index,
                )?;
                let old_thread = store.set_thread(guest_thread)?;
                log::trace!(
                    "stackless call: replaced {old_thread:?} with {guest_thread:?} as current thread"
                );

                store.enter_instance(callee_instance);

                // SAFETY: See the documentation for `make_call` to review the
                // contract we must uphold for `call` here.
                //
                // Per the contract described in the `stage_call`
                // documentation, the `callee` pointer which `call` closes
                // over must be valid.
                let storage = call(store)?;

                store.exit_instance(callee_instance)?;

                store.set_thread(old_thread)?;
                let state = store.concurrent_state_mut()?;
                if let Some(t) = old_thread.guest() {
                    state.set_thread_running(t.thread)?;
                }
                log::trace!("stackless call: restored {old_thread:?} as current thread");

                // SAFETY: `wasmparser` will have validated that the callback
                // function returns a `i32` result.
                let code = unsafe { storage[0].assume_init() }.get_i32() as u32;

                self.handle_callback_code(store, guest_thread, callee_instance.index, code)
            }) as Box<dyn FnOnce(&mut dyn VMStore) -> Result<()> + Send + Sync>
        } else {
            let token = StoreToken::new(store.as_context_mut());
            Box::new(move |store: &mut dyn VMStore| {
                self.add_guest_thread_to_instance_table(
                    guest_thread.thread,
                    store,
                    callee_instance.index,
                )?;
                let old_thread = store.set_thread(guest_thread)?;
                log::trace!(
                    "sync/async-stackful call: replaced {old_thread:?} with {guest_thread:?} as current thread",
                );
                let flags = self.id().get(store).instance_flags(callee_instance.index);

                let callee_async_typed = store
                    .concurrent_state_mut()?
                    .get_mut(guest_thread.task)?
                    .async_typed;

                // Unless this is a callback-less (i.e. stackful) async-lifted
                // or sync-typed export, we need to record that the instance
                // cannot be entered until the call returns.
                if !async_ && callee_async_typed {
                    store.enter_instance(callee_instance);
                }

                if !callee_async_typed {
                    store.enter_sync_call(callee_instance)?;
                }

                // SAFETY: See the documentation for `make_call` to review the
                // contract we must uphold for `call` here.
                //
                // Per the contract described in the `stage_call`
                // documentation, the `callee` pointer which `call` closes
                // over must be valid.
                let storage = call(store)?;

                if !callee_async_typed {
                    store.exit_sync_call(callee_instance)?;
                }

                if !async_ {
                    // This is a sync-lifted export, so now is when we lift the
                    // result, optionally call the post-return function, if any,
                    // and finally notify any current or future waiters that the
                    // subtask has returned.

                    if callee_async_typed {
                        store.exit_instance(callee_instance)?;
                    }

                    let lift = {
                        let state = store.concurrent_state_mut()?;
                        if !state.get_mut(guest_thread.task)?.result.is_none() {
                            bail_bug!("task has already produced a result");
                        }

                        match state.get_mut(guest_thread.task)?.lift_result.take() {
                            Some(lift) => lift,
                            None => bail_bug!("lift_result field is missing"),
                        }
                    };

                    // SAFETY: `result_count` represents the number of core Wasm
                    // results returned, per `wasmparser`.
                    let result = (lift.lift)(store, unsafe {
                        mem::transmute::<&[MaybeUninit<ValRaw>], &[ValRaw]>(
                            &storage[..result_count],
                        )
                    })?;

                    let post_return_arg = match result_count {
                        0 => ValRaw::i32(0),
                        // SAFETY: `result_count` represents the number of
                        // core Wasm results returned, per `wasmparser`.
                        1 => unsafe { storage[0].assume_init() },
                        _ => unreachable!(),
                    };

                    unsafe {
                        call_post_return(
                            token.as_context_mut(store),
                            post_return.map(|v| v.as_non_null()),
                            post_return_arg,
                            flags,
                        )?;
                    }

                    self.task_complete(store, guest_thread.task, result, Status::Returned)?;
                }

                store.set_thread(old_thread)?;

                store
                    .concurrent_state_mut()?
                    .get_mut(guest_thread.task)?
                    .exited = true;

                log::trace!(
                    "clean up thread; async lifted? {async_} async typed? {callee_async_typed}"
                );

                if callee_async_typed {
                    // If we're async-typed, returning control to our caller
                    // won't help resolve any outstanding sync-typed call which
                    // might be in progress, so we may need to switch or trap
                    // before exiting this thread:
                    store.switch_or_trap_if_may_not_suspend(callee_instance)?;
                }

                // This is a callback-less call, so the implicit thread has now completed
                store.cleanup_thread(guest_thread, callee_instance)?;
                Ok(())
            })
        };

        store.0.concurrent_state_mut()?.push_work_item(
            WorkItem::GuestCall {
                instance: callee_instance,
                call: GuestCall {
                    thread: guest_thread,
                    kind: GuestCallKind::StartImplicit(fun),
                },
            },
            if host_caller {
                Priority::High
            } else {
                Priority::Switch
            },
        )?;

        Ok(())
    }

    /// Prepare (but do not start) a guest->guest call.
    ///
    /// This is called from fused adapter code generated in
    /// `wasmtime_environ::fact::trampoline::Compiler`.  `start` and `return_`
    /// are synthesized Wasm functions which move the parameters from the caller
    /// to the callee and the result from the callee to the caller,
    /// respectively.  The adapter will call `Self::start_call` immediately
    /// after calling this function.
    ///
    /// SAFETY: All the pointer arguments must be valid pointers to guest
    /// entities (and with the expected signatures for the function references
    /// -- see `wasmtime_environ::fact::trampoline::Compiler` for details).
    unsafe fn prepare_call<T: 'static>(
        self,
        mut store: StoreContextMut<T>,
        start: NonNull<VMFuncRef>,
        return_: NonNull<VMFuncRef>,
        caller_instance: RuntimeComponentInstanceIndex,
        callee_instance: RuntimeComponentInstanceIndex,
        task_return_type: TypeTupleIndex,
        callee_async_typed: bool,
        memory: *mut VMMemoryDefinition,
        string_encoding: StringEncoding,
        caller_info: CallerInfo,
    ) -> Result<()> {
        enum ResultInfo {
            Heap { results: u32 },
            Stack { result_count: u32 },
        }

        let result_info = match &caller_info {
            CallerInfo::Async {
                has_result: true,
                params,
            } => ResultInfo::Heap {
                results: match params.last() {
                    Some(r) => r.get_u32(),
                    None => bail_bug!("retptr missing"),
                },
            },
            CallerInfo::Async {
                has_result: false, ..
            } => ResultInfo::Stack { result_count: 0 },
            CallerInfo::Sync {
                result_count,
                params,
            } if *result_count > u32::try_from(MAX_FLAT_RESULTS)? => ResultInfo::Heap {
                results: match params.last() {
                    Some(r) => r.get_u32(),
                    None => bail_bug!("arg ptr missing"),
                },
            },
            CallerInfo::Sync { result_count, .. } => ResultInfo::Stack {
                result_count: *result_count,
            },
        };

        let sync_caller = matches!(caller_info, CallerInfo::Sync { .. });

        // Create a new guest task for the call, closing over the `start` and
        // `return_` functions to lift the parameters and lower the result,
        // respectively.
        let start = SendSyncPtr::new(start);
        let return_ = SendSyncPtr::new(return_);
        let token = StoreToken::new(store.as_context_mut());
        let old_thread = store.0.current_guest_thread()?;

        let state = store.0.concurrent_state_mut()?;

        debug_assert_eq!(
            state.get_mut(old_thread.task)?.instance,
            self.runtime_instance(caller_instance)
        );

        let guest_thread = GuestTask::new(
            state,
            Box::new(move |store, dst| {
                let mut store = token.as_context_mut(store);
                assert!(dst.len() <= MAX_FLAT_PARAMS);
                // The `+ 1` here accounts for the return pointer, if any:
                let mut src = [MaybeUninit::uninit(); MAX_FLAT_PARAMS + 1];
                let count = match caller_info {
                    // Async callers, if they have a result, use the last
                    // parameter as a return pointer so chop that off if
                    // relevant here.
                    CallerInfo::Async { params, has_result } => {
                        let params = &params[..params.len() - usize::from(has_result)];
                        for (param, src) in params.iter().zip(&mut src) {
                            src.write(*param);
                        }
                        params.len()
                    }

                    // Sync callers forward everything directly.
                    CallerInfo::Sync { params, .. } => {
                        for (param, src) in params.iter().zip(&mut src) {
                            src.write(*param);
                        }
                        params.len()
                    }
                };
                // SAFETY: `start` is a valid `*mut VMFuncRef` from
                // `wasmtime-cranelift`-generated fused adapter code.  Based on
                // how it was constructed (see
                // `wasmtime_environ::fact::trampoline::Compiler::compile_async_start_adapter`
                // for details) we know it takes count parameters and returns
                // `dst.len()` results.
                unsafe {
                    crate::Func::call_unchecked_raw(
                        &mut store,
                        start.as_non_null(),
                        NonNull::new(
                            &mut src[..count.max(dst.len())] as *mut [MaybeUninit<ValRaw>] as _,
                        )
                        .unwrap(),
                        UncaughtException::Trap,
                    )?;
                }
                dst.copy_from_slice(&src[..dst.len()]);
                let task = store.0.current_guest_thread()?.task;
                let state = store.0.concurrent_state_mut()?;
                Waitable::Guest(task).set_event(
                    state,
                    Some(Event::Subtask {
                        status: Status::Started,
                    }),
                )?;
                Ok(())
            }),
            LiftResult {
                lift: Box::new(move |store, src| {
                    // SAFETY: See comment in closure passed as `lower_params`
                    // parameter above.
                    let mut store = token.as_context_mut(store);
                    let mut my_src = src.to_owned(); // TODO: use stack to avoid allocation?
                    if let ResultInfo::Heap { results } = &result_info {
                        my_src.push(ValRaw::u32(*results));
                    }

                    // SAFETY: `return_` is a valid `*mut VMFuncRef` from
                    // `wasmtime-cranelift`-generated fused adapter code.  Based
                    // on how it was constructed (see
                    // `wasmtime_environ::fact::trampoline::Compiler::compile_async_return_adapter`
                    // for details) we know it takes `src.len()` parameters and
                    // returns up to 1 result.
                    unsafe {
                        crate::Func::call_unchecked_raw(
                            &mut store,
                            return_.as_non_null(),
                            my_src.as_mut_slice().into(),
                            UncaughtException::Trap,
                        )?;
                    }

                    let thread = store.0.current_guest_thread()?;
                    let state = store.0.concurrent_state_mut()?;
                    if sync_caller {
                        state.get_mut(thread.task)?.sync_result = SyncResult::Produced(
                            if let ResultInfo::Stack { result_count } = &result_info {
                                match result_count {
                                    0 => None,
                                    1 => Some(my_src[0]),
                                    _ => unreachable!(),
                                }
                            } else {
                                None
                            },
                        );
                    }
                    Ok(Box::new(DummyResult) as Box<dyn Any + Send + Sync>)
                }),
                ty: task_return_type,
                memory: NonNull::new(memory).map(SendSyncPtr::new),
                string_encoding,
            },
            Caller::Guest { thread: old_thread },
            None,
            self.runtime_instance(callee_instance),
            callee_async_typed,
            // We don't know whether the callee export was lifted sync or async
            // yet, but we'll update this in `start_call`:
            false,
        )?;

        // Make the new thread the current one so that `Self::start_call` knows
        // which one to start.
        store.0.set_thread(guest_thread)?;
        log::trace!("pushed {guest_thread:?} as current thread; old thread was {old_thread:?}");

        Ok(())
    }

    /// Call the specified callback function for an async-lifted export.
    ///
    /// SAFETY: `function` must be a valid reference to a guest function of the
    /// correct signature for a callback.
    unsafe fn call_callback<T>(
        self,
        mut store: StoreContextMut<T>,
        function: SendSyncPtr<VMFuncRef>,
        event: Event,
        handle: u32,
    ) -> Result<u32> {
        let (ordinal, result) = event.parts();
        let params = &mut [
            ValRaw::u32(ordinal),
            ValRaw::u32(handle),
            ValRaw::u32(result),
        ];
        // SAFETY: `func` is a valid `*mut VMFuncRef` from either
        // `wasmtime-cranelift`-generated fused adapter code or
        // `component::Options`.  Per `wasmparser` callback signature
        // validation, we know it takes three parameters and returns one.
        unsafe {
            crate::Func::call_unchecked_raw(
                &mut store,
                function.as_non_null(),
                params.as_mut_slice().into(),
                UncaughtException::Trap,
            )?;
        }
        Ok(params[0].get_u32())
    }

    /// Start a guest->guest call previously prepared using
    /// `Self::prepare_call`.
    ///
    /// This is called from fused adapter code generated in
    /// `wasmtime_environ::fact::trampoline::Compiler`.  The adapter will call
    /// this function immediately after calling `Self::prepare_call`.
    ///
    /// If `flags` contains `START_FLAG_ASYNC_CALLER` then the caller used an
    /// async-lowered import and the call's packed status is written to
    /// `storage[0]`. Otherwise this blocks until the callee has produced its
    /// result, if any, which is written to `storage[0]`.
    ///
    /// SAFETY: The `*mut VMFuncRef` arguments must be valid pointers to guest
    /// functions with the appropriate signatures for the current guest task.
    /// If this is a call to an async-lowered import, the actual call may be
    /// deferred and run after this function returns, in which case the pointer
    /// arguments must also be valid when the call happens.
    unsafe fn start_call<T: 'static>(
        self,
        mut store: StoreContextMut<T>,
        callback: *mut VMFuncRef,
        post_return: *mut VMFuncRef,
        callee: NonNull<VMFuncRef>,
        param_count: u32,
        result_count: u32,
        flags: u32,
        storage: &mut [MaybeUninit<ValRaw>],
    ) -> Result<()> {
        let token = StoreToken::new(store.as_context_mut());
        let async_caller = (flags & START_FLAG_ASYNC_CALLER) != 0;
        let guest_thread = store.0.current_guest_thread()?;
        let state = store.0.concurrent_state_mut()?;

        if !state.event_loop_running {
            bail_bug!("Instance::start_call called without a running event loop");
        }

        let callee = SendSyncPtr::new(callee);
        let param_count = usize::try_from(param_count)?;
        assert!(param_count <= MAX_FLAT_PARAMS);
        let result_count = usize::try_from(result_count)?;
        assert!(result_count <= MAX_FLAT_RESULTS);

        let task = state.get_mut(guest_thread.task)?;
        let callee_async_typed = task.async_typed;
        let callee_instance = task.instance;

        task.async_lifted = (flags & START_FLAG_ASYNC_CALLEE) != 0;

        if let Some(callback) = NonNull::new(callback) {
            // We're calling an async-lifted export with a callback, so store
            // the callback and related context as part of the task so we can
            // call it later when needed.
            let callback = SendSyncPtr::new(callback);
            task.callback = Some(Box::new(move |store, event, handle| {
                let store = token.as_context_mut(store);
                unsafe { self.call_callback::<T>(store, callback, event, handle) }
            }));
        }

        let Caller::Guest { thread: caller } = &task.caller else {
            // As of this writing, `start_call` is only used for guest->guest
            // calls.
            bail_bug!("start_call unexpectedly invoked for host->guest call");
        };
        let caller = *caller;
        let caller_instance = state.get_mut(caller.task)?.instance;

        // Stage the call as the work item to run next, modulo backpressure, etc.
        unsafe {
            self.stage_call(
                store.as_context_mut(),
                guest_thread,
                callee,
                param_count,
                result_count,
                (flags & START_FLAG_ASYNC_CALLEE) != 0,
                NonNull::new(callback).map(SendSyncPtr::new),
                NonNull::new(post_return).map(SendSyncPtr::new),
                false,
            )?;
        }

        let old_do_not_suspend = if callee_async_typed {
            // If we're starting an async-typed call, it is permitted to suspend
            // since that will just return control back to the caller, which may
            // be able to avoid blocking if needed.
            //
            // We'll restore the old value after the call either returns or
            // suspends.
            let state = store.0.instance_state(callee_instance).concurrent_state();
            let old_do_not_suspend = state.do_not_suspend;
            state.do_not_suspend = false;
            Some(old_do_not_suspend)
        } else {
            None
        };

        let state = store.0.concurrent_state_mut()?;

        // Use the caller's `GuestThread::sync_call_set` to register interest in
        // the subtask...
        let guest_waitable = Waitable::Guest(guest_thread.task);
        let old_set = guest_waitable.common(state)?.set;
        let set = state.get_mut(caller.thread)?.sync_call_set;
        guest_waitable.join(state, Some(set))?;

        store.0.set_thread(CurrentThread::None)?;

        // ... and suspend this fiber temporarily while we wait for it to start.
        //
        // Note that we _could_ call the callee directly using the current fiber
        // rather than suspend this one, but that would make reasoning about the
        // event loop more complicated and is probably only worth doing if
        // there's a measurable performance benefit.  In addition, it would mean
        // blocking the caller if the callee calls a blocking sync-lowered
        // import, and as of this writing the spec says we must not do that.
        //
        // Alternatively, the fused adapter code could be modified to call the
        // callee directly without calling a host-provided intrinsic at all (in
        // which case it would need to do its own, inline backpressure checks,
        // etc.).  Again, we'd want to see a measurable performance benefit
        // before committing to such an optimization.  And again, we'd need to
        // update the spec to allow that.
        let mut yielded = false;
        let (status, waitable) = loop {
            store.0.suspend(if yielded {
                SuspendReason::Waiting {
                    set,
                    thread: caller,
                }
            } else {
                yielded = true;
                SuspendReason::YieldingToSubtask { thread: caller }
            })?;

            if let Some(old_do_not_suspend) = old_do_not_suspend {
                store
                    .0
                    .instance_state(callee_instance)
                    .concurrent_state()
                    .do_not_suspend = old_do_not_suspend;
            }

            let state = store.0.concurrent_state_mut()?;

            log::trace!("taking event for {:?}", guest_thread.task);
            let event = guest_waitable.take_event(state)?;
            let Some(Event::Subtask { status }) = event else {
                bail_bug!("subtasks should only get subtask events, got {event:?}")
            };

            log::trace!("status {status:?} for {:?}", guest_thread.task);

            if status == Status::Returned {
                // It returned, so we can stop waiting.
                break (status, None);
            } else if async_caller {
                // It hasn't returned yet, but the caller is calling via an
                // async-lowered import, so we generate a handle for the task
                // waitable and return the status.
                let handle = store
                    .0
                    .instance_state(caller_instance)
                    .handle_table()
                    .subtask_insert_guest(guest_thread.task.rep())?;
                store
                    .0
                    .concurrent_state_mut()?
                    .get_mut(guest_thread.task)?
                    .common
                    .handle = Some(handle);
                break (status, Some(handle));
            } else {
                // The callee hasn't returned yet, and the caller is calling via
                // a sync-lowered import, so we loop and keep waiting until the
                // callee returns.
                store.0.switch_or_trap_if_may_not_suspend(caller_instance)?;
            }
        };

        guest_waitable.join(store.0.concurrent_state_mut()?, old_set)?;

        // Reset the current thread to point to the caller as it resumes control.
        store.0.set_thread(caller)?;

        store
            .0
            .concurrent_state_mut()?
            .set_thread_running(caller.thread)?;

        log::trace!("popped current thread {guest_thread:?}; new thread is {caller:?}");

        if async_caller {
            let Some(slot) = storage.first_mut() else {
                bail_bug!("no storage for async call status");
            };
            *slot = MaybeUninit::new(ValRaw::u32(status.pack(waitable)));
        } else {
            // The caller used a sync-lowered import, in which case the result,
            // if any, has been stashed in `GuestTask::sync_result`.
            let state = store.0.concurrent_state_mut()?;
            let task = state.get_mut(guest_thread.task)?;
            if let Some(Some(result)) = task.sync_result.take()? {
                storage[0] = MaybeUninit::new(result);
            }
        }

        if status == Status::Returned {
            store
                .0
                .concurrent_state_mut()?
                .decrement_task_ref_count(guest_thread.task)?;
        }

        Ok(())
    }

    /// Poll the specified future once on behalf of a guest->host call using an
    /// async-lowered import.
    ///
    /// If it returns `Ready`, return `Ok(None)`.  Otherwise, if it returns
    /// `Pending`, add it to the set of futures to be polled as part of this
    /// instance's event loop until it completes, and then return
    /// `Ok(Some(handle))` where `handle` is the waitable handle to return.
    ///
    /// Whether the future returns `Ready` immediately or later, the `lower`
    /// function will be used to lower the result, if any, into the guest caller's
    /// stack and linear memory. The `lower` function is invoked with the
    /// `Option<R>` param `None` if the future is cancelled. The
    /// `Option<TableId<HostTask>>` is passed as `Some` if the host task was
    /// materialized during execution and allows `lower` to delete the task if
    /// needed.
    pub(crate) fn first_poll<T: 'static, R: Send + 'static>(
        self,
        mut store: StoreContextMut<'_, T>,
        host_task: EnteredHostTask,
        result_may_require_realloc: bool,
        future: impl Future<Output = Result<R>> + Send + 'static,
        lower: impl FnOnce(StoreContextMut<T>, Option<R>, bool, Option<TableId<HostTask>>) -> Result<()>
        + Send
        + 'static,
    ) -> Result<u32> {
        let token = StoreToken::new(store.as_context_mut());

        // Create an abortable future which hooks calls to poll and manages call
        // context state for the future.
        let (join_handle, future) = JoinHandle::run(future);
        let mut future = Box::pin(future);

        // Finally, poll the future.  We can use a dummy `Waker` here because
        // we'll add the future to `ConcurrentState::futures` and poll it
        // automatically from the event loop if it doesn't complete immediately
        // here.
        let poll = tls::set(store.0, || {
            future
                .as_mut()
                .poll(&mut Context::from_waker(&Waker::noop()))
        });

        match poll {
            // It finished immediately; lower the result and delete the task.
            Poll::Ready(result) => {
                let result = result.transpose()?;
                // Check if the host task was materialized so that it can be
                // deleted in `lower`.
                let task = store.0.current_materialized_host_task()?;
                lower(store.as_context_mut(), result, true, task)?;
                return Ok(Status::Returned.pack(None));
            }

            // Future isn't ready yet, so fall through.
            Poll::Pending => {}
        }

        // The future will outlive this call frame, so materialize the deferred
        // host task and attach its cancellation handle before publishing it to
        // the event loop.
        let Some(task) = store.0.materialize_host_task_id()? else {
            bail_bug!("current thread is not a host thread")
        };
        {
            let state = &mut store.0.concurrent_state_mut()?.get_mut(task)?.state;
            assert!(matches!(state, HostTaskState::CalleeStarted));
            *state = HostTaskState::CalleeRunning(join_handle);
        }

        // It hasn't finished yet; add the future to
        // `ConcurrentState::futures` so it will be polled by the event
        // loop and allocate a waitable handle to return to the guest.

        // Wrap the future in a closure responsible for lowering the result into
        // the guest's stack and memory, as well as notifying any waiters that
        // the task returned.
        let future = Box::pin(async move {
            let result = match run_with_host_task_set(task, future).await? {
                Some(result) => Some(result?),
                None => None,
            };
            let on_complete = move |store: &mut dyn VMStore| {
                // Restore the `current_thread` to be the host so `lower` knows
                // how to manipulate borrows and knows which scope of borrows
                // to check.
                let mut store = token.as_context_mut(store);
                let old = store.0.set_thread(task)?;

                let status = if result.is_some() {
                    Status::Returned
                } else {
                    Status::ReturnCancelled
                };

                lower(store.as_context_mut(), result, false, Some(task))?;
                let state = store.0.concurrent_state_mut()?;
                match &mut state.get_mut(task)?.state {
                    // Cancellation is only complete once lowering has finished
                    // and the terminal event can be published.
                    pending @ HostTaskState::CalleeCancelling => {
                        *pending = HostTaskState::CalleeDone { cancelled: true };
                    }

                    // Otherwise transition this task to the done state.
                    other => *other = HostTaskState::CalleeDone { cancelled: false },
                }
                Waitable::Host(task).set_event(state, Some(Event::Subtask { status }))?;

                store.0.set_thread(old)?;
                Ok(())
            };

            tls::get(move |store| {
                if result_may_require_realloc {
                    // Lowering may call the guest's realloc function, so run
                    // it on a worker fiber. This is required because the guest
                    // might try to suspend the fiber which is unsound when it's
                    // not actually on a fiber.
                    store
                        .concurrent_state_mut()?
                        .push_high_priority(WorkItem::WorkerFunction(AlwaysMut::new(Box::new(
                            on_complete,
                        ))));
                    Ok(())
                } else {
                    // No guest code can be called while lowering this result,
                    // so avoid scheduling a worker fiber.
                    on_complete(store)
                }
            })
        });

        // Make this task visible to the guest and then record what it
        // was made visible as.
        let caller = match host_task {
            Some(caller) => caller,
            None => bail_bug!("host task wasn't created but should have been"),
        };
        let state = store.0.concurrent_state_mut()?;
        state.push_future(future);
        let instance = state.get_mut(caller.task)?.instance;
        let handle = store
            .0
            .instance_state(instance)
            .handle_table()
            .subtask_insert_host(task.rep())?;
        store.0.concurrent_state_mut()?.get_mut(task)?.common.handle = Some(handle);
        log::trace!("assign {task:?} handle {handle} for {caller:?} instance {instance:?}");

        // Restore the currently running thread to this host task's
        // caller. Note that the host task isn't deallocated as it's
        // within the store and will get deallocated later.
        store.0.set_thread(caller)?;
        Ok(Status::Started.pack(Some(handle)))
    }

    /// Implements the `task.return` intrinsic, lifting the result for the
    /// current guest task.
    pub(crate) fn task_return(
        self,
        store: &mut dyn VMStore,
        ty: TypeTupleIndex,
        options: OptionsIndex,
        storage: &[ValRaw],
    ) -> Result<()> {
        let guest_thread = store.current_guest_thread()?;
        let state = store.concurrent_state_mut()?;
        if !state.get_mut(guest_thread.task)?.async_lifted {
            bail!(Trap::TaskReturnOrCancelSyncLifted);
        }
        let lift = state
            .get_mut(guest_thread.task)?
            .lift_result
            .take()
            .ok_or_else(|| Trap::TaskCancelOrReturnTwice)?;
        if !state.get_mut(guest_thread.task)?.result.is_none() {
            bail_bug!("task result unexpectedly already set");
        }

        let CanonicalOptions {
            string_encoding,
            data_model,
            ..
        } = &self.id().get(store).component().env_component().options[options];

        let invalid = ty != lift.ty
            || string_encoding != &lift.string_encoding
            || match data_model {
                CanonicalOptionsDataModel::LinearMemory(opts) => match opts.memory {
                    Some(memory) => {
                        let expected = lift.memory.map(|v| v.as_ptr()).unwrap_or(ptr::null_mut());
                        let actual = self.id().get(store).runtime_memory(memory);
                        expected != actual.as_ptr()
                    }
                    // Memory not specified, meaning it didn't need to be
                    // specified per validation, so not invalid.
                    None => false,
                },
                // Always invalid as this isn't supported.
                CanonicalOptionsDataModel::Gc { .. } => true,
            };

        if invalid {
            bail!(Trap::TaskReturnInvalid);
        }

        log::trace!("task.return for {guest_thread:?}");

        let result = (lift.lift)(store, storage)?;
        self.task_complete(store, guest_thread.task, result, Status::Returned)
    }

    /// Implements the `task.cancel` intrinsic.
    pub(crate) fn task_cancel(self, store: &mut StoreOpaque) -> Result<()> {
        let guest_thread = store.current_guest_thread()?;
        let state = store.concurrent_state_mut()?;
        let task = state.get_mut(guest_thread.task)?;
        if !task.async_lifted {
            bail!(Trap::TaskReturnOrCancelSyncLifted);
        }
        if !task.cancel_request_delivered {
            bail!(Trap::TaskCancelNotCancelled);
        }
        _ = task
            .lift_result
            .take()
            .ok_or_else(|| Trap::TaskCancelOrReturnTwice)?;

        if !task.result.is_none() {
            bail_bug!("task result should not bet set yet");
        }

        log::trace!("task.cancel for {guest_thread:?}");

        self.task_complete(
            store,
            guest_thread.task,
            Box::new(DummyResult),
            Status::ReturnCancelled,
        )
    }

    /// Complete the specified guest task (i.e. indicate that it has either
    /// returned a (possibly empty) result or cancelled itself).
    ///
    /// This will return any resource borrows and notify any current or future
    /// waiters that the task has completed.
    fn task_complete(
        self,
        store: &mut StoreOpaque,
        guest_task: TableId<GuestTask>,
        result: Box<dyn Any + Send + Sync>,
        status: Status,
    ) -> Result<()> {
        store
            .component_resource_tables(Some(self))?
            .validate_scope_exit()?;

        let state = store.concurrent_state_mut()?;
        let task = state.get_mut(guest_task)?;

        // A cancellation request is only delivered to a task which hasn't yet
        // resolved, so drop any pending, undelivered `Event::Cancelled` now
        // that this task has resolved.
        task.event = None;

        if let Caller::Host { tx, .. } = &mut task.caller {
            if let Some(tx) = tx.take() {
                _ = tx.send(result);
            }
        } else {
            task.result = Some(result);
            Waitable::Guest(guest_task).set_event(state, Some(Event::Subtask { status }))?;
        }

        Ok(())
    }

    /// Implements the `waitable-set.new` intrinsic.
    pub(crate) fn waitable_set_new(
        self,
        store: &mut StoreOpaque,
        caller_instance: RuntimeComponentInstanceIndex,
    ) -> Result<u32> {
        let set = store.concurrent_state_mut()?.push(WaitableSet::default())?;
        let handle = store
            .instance_state(self.runtime_instance(caller_instance))
            .handle_table()
            .waitable_set_insert(set.rep())?;
        log::trace!("new waitable set {set:?} (handle {handle})");
        Ok(handle)
    }

    /// Implements the `waitable-set.drop` intrinsic.
    pub(crate) fn waitable_set_drop(
        self,
        store: &mut StoreOpaque,
        caller_instance: RuntimeComponentInstanceIndex,
        set: u32,
    ) -> Result<()> {
        let rep = store
            .instance_state(self.runtime_instance(caller_instance))
            .handle_table()
            .waitable_set_remove(set)?;

        log::trace!("drop waitable set {rep} (handle {set})");

        // Note that we're careful to check for waiters _before_ deleting the
        // set to avoid dropping any waiters in `WaitMode::Fiber(_)`, which
        // would panic.  See `drop-waitable-set-with-waiters.wast` for details.
        let set = store
            .concurrent_state_mut()?
            .get_mut(TableId::<WaitableSet>::new(rep))?;
        if set.num_waiting > 0 {
            bail!(Trap::WaitableSetDropHasWaiters);
        }

        store
            .concurrent_state_mut()?
            .delete(TableId::<WaitableSet>::new(rep))?;

        Ok(())
    }

    /// Implements the `waitable.join` intrinsic.
    pub(crate) fn waitable_join(
        self,
        store: &mut StoreOpaque,
        caller_instance: RuntimeComponentInstanceIndex,
        waitable_handle: u32,
        set_handle: u32,
    ) -> Result<()> {
        let mut instance = self.id().get_mut(store);
        let waitable =
            Waitable::from_instance(instance.as_mut(), caller_instance, waitable_handle)?;

        let set = if set_handle == 0 {
            None
        } else {
            let set = instance.instance_states().0[caller_instance]
                .handle_table()
                .waitable_set_rep(set_handle)?;

            Some(TableId::<WaitableSet>::new(set))
        };

        waitable.trap_if_has_sync_waiter(store.concurrent_state_mut()?)?;

        log::trace!(
            "waitable {waitable:?} (handle {waitable_handle}) join set {set:?} (handle {set_handle})",
        );

        waitable.join(store.concurrent_state_mut()?, set)
    }

    /// Implements the `subtask.drop` intrinsic.
    pub(crate) fn subtask_drop(
        self,
        store: &mut StoreOpaque,
        caller_instance: RuntimeComponentInstanceIndex,
        task_id: u32,
    ) -> Result<()> {
        self.waitable_join(store, caller_instance, task_id, 0)?;

        let (rep, is_host) = store
            .instance_state(self.runtime_instance(caller_instance))
            .handle_table()
            .subtask_remove(task_id)?;

        let concurrent_state = store.concurrent_state_mut()?;
        let waitable = if is_host {
            let id = TableId::<HostTask>::new(rep);
            let task = concurrent_state.get_mut(id)?;
            match &task.state {
                HostTaskState::CalleeRunning(_) | HostTaskState::CalleeCancelling => {
                    bail!(Trap::SubtaskDropNotResolved)
                }
                HostTaskState::CalleeDone { .. } => {}
                HostTaskState::CalleeStarted | HostTaskState::CalleeFinished(_) => {
                    bail_bug!("invalid state for callee in `subtask.drop`")
                }
            }

            Waitable::Host(id)
        } else {
            let id = TableId::<GuestTask>::new(rep);
            let task = concurrent_state.get_mut(id)?;
            if task.lift_result.is_some() {
                bail!(Trap::SubtaskDropNotResolved);
            }
            Waitable::Guest(id)
        };

        waitable.common(concurrent_state)?.handle = None;

        // If this subtask has an event that means that the terminal status of
        // this subtask wasn't yet received so it can't be dropped yet.
        if waitable.take_event(concurrent_state)?.is_some() {
            bail!(Trap::SubtaskDropNotResolved);
        }

        waitable.delete_from(concurrent_state)?;

        log::trace!("subtask_drop {waitable:?} (handle {task_id})");
        Ok(())
    }

    /// Implements the `waitable-set.wait` intrinsic.
    pub(crate) fn waitable_set_wait(
        self,
        store: &mut StoreOpaque,
        options: OptionsIndex,
        set: u32,
        payload: u32,
    ) -> Result<u32> {
        let &CanonicalOptions {
            instance: caller_instance,
            ..
        } = &self.id().get(store).component().env_component().options[options];
        let caller = self.runtime_instance(caller_instance);
        let rep = store
            .instance_state(self.runtime_instance(caller_instance))
            .handle_table()
            .waitable_set_rep(set)?;

        self.waitable_check(
            store,
            caller,
            WaitableCheck::Wait,
            WaitableCheckParams {
                set: TableId::new(rep),
                options,
                payload,
            },
        )
    }

    /// Implements the `waitable-set.poll` intrinsic.
    pub(crate) fn waitable_set_poll(
        self,
        store: &mut StoreOpaque,
        options: OptionsIndex,
        set: u32,
        payload: u32,
    ) -> Result<u32> {
        let &CanonicalOptions {
            instance: caller_instance,
            ..
        } = &self.id().get(store).component().env_component().options[options];
        let caller = self.runtime_instance(caller_instance);
        let rep = store
            .instance_state(caller)
            .handle_table()
            .waitable_set_rep(set)?;

        self.waitable_check(
            store,
            caller,
            WaitableCheck::Poll,
            WaitableCheckParams {
                set: TableId::new(rep),
                options,
                payload,
            },
        )
    }

    /// Implements the `thread.index` intrinsic.
    pub(crate) fn thread_index(&self, store: &mut dyn VMStore) -> Result<u32> {
        let thread_id = store.current_guest_thread()?.thread;
        match store
            .concurrent_state_mut()?
            .get_mut(thread_id)?
            .instance_rep
        {
            Some(r) => Ok(r),
            None => bail_bug!("thread should have instance_rep by now"),
        }
    }

    /// Implements the `thread.new-indirect` intrinsic.
    pub(crate) fn thread_new_indirect<T: 'static>(
        self,
        mut store: StoreContextMut<T>,
        runtime_instance: RuntimeComponentInstanceIndex,
        start_func_ty: ModuleInternedTypeIndex,
        start_func_table_idx: RuntimeTableIndex,
        start_func_idx: u32,
        context: i32,
    ) -> Result<u32> {
        log::trace!("creating new thread");

        let (instance, registry) = self.id().get_mut_and_registry(store.0);
        let Some(start_func_ty) = instance.component().signatures().shared_type(start_func_ty)
        else {
            bail_bug!("thread.new-indirect start function type should be registered");
        };
        let callee = instance
            .index_runtime_func_table(registry, start_func_table_idx, start_func_idx as u64)?
            .ok_or_else(|| Trap::ThreadNewIndirectUninitialized)?;
        if callee.type_index(store.0) != start_func_ty {
            bail!(Trap::ThreadNewIndirectInvalidType);
        }

        let token = StoreToken::new(store.as_context_mut());
        let start_func = Box::new(
            move |store: &mut dyn VMStore, guest_thread: QualifiedThreadId| -> Result<()> {
                let old_thread = store.set_thread(guest_thread)?;
                log::trace!(
                    "thread start: replaced {old_thread:?} with {guest_thread:?} as current thread"
                );

                let mut store = token.as_context_mut(store);
                let mut params = [ValRaw::i32(context)];
                // Use call_unchecked rather than call or call_async, as we don't want to run the function
                // on a separate fiber if we're running in an async store.
                unsafe { callee.call_unchecked(store.as_context_mut(), &mut params)? };

                store.0.set_thread(old_thread)?;

                let runtime_instance = self.runtime_instance(runtime_instance);

                // We're not returning to any caller, so we may need to switch
                // or trap if the instance has an outstanding sync-typed call:
                store
                    .0
                    .switch_or_trap_if_may_not_suspend(runtime_instance)?;

                store.0.cleanup_thread(guest_thread, runtime_instance)?;

                log::trace!("explicit thread {guest_thread:?} completed");
                let state = store.0.concurrent_state_mut()?;
                if let Some(t) = old_thread.guest() {
                    state.set_thread_running(t.thread)?;
                }
                log::trace!("thread start: restored {old_thread:?} as current thread");

                Ok(())
            },
        );

        let current_thread = store.0.current_guest_thread()?;
        let state = store.0.concurrent_state_mut()?;
        let parent_task = current_thread.task;

        let new_thread = GuestThread::new_explicit(state, parent_task, start_func)?;
        let thread_id = state.push(new_thread)?;
        state.get_mut(parent_task)?.threads.insert(thread_id);

        log::trace!("new thread with id {thread_id:?} created");

        self.add_guest_thread_to_instance_table(thread_id, store.0, runtime_instance)
    }

    pub(crate) fn resume_thread(
        self,
        store: &mut StoreOpaque,
        runtime_instance: RuntimeComponentInstanceIndex,
        thread_idx: u32,
        how: ResumeThread,
    ) -> Result<bool> {
        let thread_id =
            GuestThread::from_instance(self.id().get_mut(store), runtime_instance, thread_idx)?;
        let state = store.concurrent_state_mut()?;
        let guest_thread = QualifiedThreadId::qualify(state, thread_id)?;

        if store.current_guest_thread()? == guest_thread {
            bail!(Trap::CannotResumeThread);
        }

        let priority = match how {
            ResumeThread::Resume => Priority::Switch,
            ResumeThread::ResumeLater => Priority::Low,

            // Promotion is a noop unless the thread is ready, so only search
            // through pending work items to see if something matches.
            ResumeThread::Promote => {
                let instance = self.runtime_instance(runtime_instance);
                let do_not_enter = store
                    .instance_state(instance)
                    .concurrent_state()
                    .do_not_enter;
                return store.concurrent_state_mut()?.promote_work_item_matching(
                    |item: &WorkItem| match item {
                        WorkItem::ResumeThread { thread, .. }
                        | WorkItem::ResumeFiber { thread, .. } => *thread == guest_thread,

                        WorkItem::GuestCall {
                            call: GuestCall { thread, kind },
                            ..
                        } => {
                            *thread == guest_thread
                                && match kind {
                                    // Events can only be delivered if this
                                    // instance's exclusive lock isn't held, so
                                    // gate on that.
                                    GuestCallKind::DeliverEvent { .. } => !do_not_enter,

                                    // A not-yet-started thread made ready via
                                    // `thread.resume-later`.
                                    GuestCallKind::StartExplicit(_) => true,

                                    // Technically shouldn't be possible since
                                    // if this hasn't started nothing could get
                                    // its id...
                                    GuestCallKind::StartImplicit(_) => false,
                                }
                        }

                        WorkItem::PushFuture(_) | WorkItem::WorkerFunction(_) => false,
                    },
                );
            }
        };

        let state = store.concurrent_state_mut()?;
        let thread = state.get_mut(guest_thread.thread)?;

        // When resuming a thread it must be in a suspended state otherwise
        // this operation is a trap.
        match &thread.state {
            GuestThreadState::NotStartedExplicit(_) => {}
            GuestThreadState::Suspended(_) => {
                log::trace!("resuming thread {thread_id:?} that was suspended");
                state.schedule_suspended_thread(
                    guest_thread,
                    self.runtime_instance(runtime_instance),
                    priority,
                )?;
                return Ok(true);
            }
            _ => bail!(Trap::CannotResumeThread),
        }
        match mem::replace(&mut thread.state, GuestThreadState::Running) {
            GuestThreadState::NotStartedExplicit(start_func) => {
                log::trace!("starting thread {guest_thread:?}");
                let guest_call = WorkItem::GuestCall {
                    instance: self.runtime_instance(runtime_instance),
                    call: GuestCall {
                        thread: guest_thread,
                        kind: GuestCallKind::StartExplicit(Box::new(move |store| {
                            start_func(store, guest_thread)
                        })),
                    },
                };
                state.push_work_item(guest_call, priority)?;
            }
            other => {
                thread.state = other;
                bail_bug!("thread state checked to be resumable above");
            }
        }
        Ok(true)
    }

    fn add_guest_thread_to_instance_table(
        self,
        thread_id: TableId<GuestThread>,
        store: &mut StoreOpaque,
        runtime_instance: RuntimeComponentInstanceIndex,
    ) -> Result<u32> {
        let guest_id = store
            .instance_state(self.runtime_instance(runtime_instance))
            .thread_handle_table()
            .guest_thread_insert(thread_id.rep())?;
        store
            .concurrent_state_mut()?
            .get_mut(thread_id)?
            .instance_rep = Some(guest_id);
        Ok(guest_id)
    }

    /// Helper function for the `thread.yield`, thread.suspend`,
    /// `thread.suspend-then-resume`, `thread.suspend-then-promote`,
    /// `thread.yield-then-resume`, and `thread.yield-then-promote` intrinsics.
    pub(crate) fn suspension_intrinsic(
        self,
        store: &mut StoreOpaque,
        caller: RuntimeComponentInstanceIndex,
        yielding: bool,
        to_thread: SuspensionTarget,
    ) -> Result<WaitResult> {
        let check_suspend = match to_thread {
            SuspensionTarget::Promote(thread) => {
                !self.resume_thread(store, caller, thread, ResumeThread::Promote)?
            }
            SuspensionTarget::Resume(thread) => {
                if !self.resume_thread(store, caller, thread, ResumeThread::Resume)? {
                    bail_bug!(
                        "`resume_thread` should only ever return false \
                         when `ResumeThread::Promote` is passed to it"
                    );
                }
                false
            }
            SuspensionTarget::None => true,
        };

        if check_suspend && !store.switch_if_may_not_suspend(self.runtime_instance(caller))? {
            return if yielding {
                Ok(WaitResult::Completed)
            } else {
                Err(Trap::CannotBlockSyncTask.into())
            };
        }

        let guest_thread = store.current_guest_thread()?;

        let reason = if yielding {
            SuspendReason::Yielding {
                thread: guest_thread,
            }
        } else {
            SuspendReason::ExplicitlySuspending {
                thread: guest_thread,
            }
        };

        store.suspend(reason)?;

        Ok(WaitResult::Completed)
    }

    /// Helper function for the `waitable-set.wait` and `waitable-set.poll` intrinsics.
    fn waitable_check(
        self,
        store: &mut StoreOpaque,
        caller: RuntimeInstance,
        check: WaitableCheck,
        params: WaitableCheckParams,
    ) -> Result<u32> {
        let guest_thread = store.current_guest_thread()?;

        log::trace!("waitable check for {guest_thread:?}; set {:?}", params.set);

        // If we're waiting, and there are no events immediately available,
        // suspend the fiber until that changes.
        match &check {
            WaitableCheck::Wait => {
                let set = params.set;

                // Note that this is a loop since, by the time this fiber is
                // resumed after being woken up, another thread may have already
                // taken the event (e.g. via `waitable-set.poll`), in which case
                // we keep waiting.
                loop {
                    let state = store.concurrent_state_mut()?;
                    let task = state.get_mut(guest_thread.task)?;
                    if !(task.event.is_none() || matches!(task.event, Some(Event::Cancelled)))
                        || !state.get_mut(set)?.ready.is_empty()
                    {
                        break;
                    }

                    store.switch_or_trap_if_may_not_suspend(caller)?;

                    store.suspend(SuspendReason::Waiting {
                        set,
                        thread: guest_thread,
                    })?;
                }
            }
            WaitableCheck::Poll => {}
        }

        log::trace!(
            "waitable check for {guest_thread:?}; set {:?}, part two",
            params.set
        );

        // Deliver any pending events to the guest and return.
        let event = self.get_event(store, guest_thread.task, Some(params.set), false)?;

        let (ordinal, handle, result) = match &check {
            WaitableCheck::Wait => {
                let (event, waitable) = match event {
                    Some(p) => p,
                    None => bail_bug!("event expected to be present"),
                };
                let handle = waitable.map(|(_, v)| v).unwrap_or(0);
                let (ordinal, result) = event.parts();
                (ordinal, handle, result)
            }
            WaitableCheck::Poll => {
                if let Some((event, waitable)) = event {
                    let handle = waitable.map(|(_, v)| v).unwrap_or(0);
                    let (ordinal, result) = event.parts();
                    (ordinal, handle, result)
                } else {
                    log::trace!(
                        "no events ready to deliver via waitable-set.poll to {:?}; set {:?}",
                        guest_thread.task,
                        params.set
                    );
                    let (ordinal, result) = Event::None.parts();
                    (ordinal, 0, result)
                }
            }
        };
        let memory = self.options_memory_mut(store, params.options);
        let ptr = crate::component::func::validate_inbounds_dynamic(
            &CanonicalAbiInfo::POINTER_PAIR,
            memory,
            &ValRaw::u32(params.payload),
        )?;
        memory[ptr + 0..][..4].copy_from_slice(&handle.to_le_bytes());
        memory[ptr + 4..][..4].copy_from_slice(&result.to_le_bytes());
        Ok(ordinal)
    }

    /// Implements the `subtask.cancel` intrinsic.
    pub(crate) fn subtask_cancel(
        self,
        store: &mut StoreOpaque,
        caller_instance: RuntimeComponentInstanceIndex,
        async_: bool,
        task_id: u32,
    ) -> Result<u32> {
        let (rep, is_host) = store
            .instance_state(self.runtime_instance(caller_instance))
            .handle_table()
            .subtask_rep(task_id)?;
        let waitable = if is_host {
            Waitable::Host(TableId::<HostTask>::new(rep))
        } else {
            Waitable::Guest(TableId::<GuestTask>::new(rep))
        };
        let concurrent_state = store.concurrent_state_mut()?;

        log::trace!("subtask_cancel {waitable:?} (handle {task_id}; async {async_})");

        waitable.trap_if_in_waitable_set(concurrent_state)?;

        let needs_block;
        if let Waitable::Host(host_task) = waitable {
            let state = &mut concurrent_state.get_mut(host_task)?.state;
            match state {
                // If the callee is still running, signal an abort is requested.
                //
                // After cancelling this falls through to block waiting for the
                // host task to actually finish assuming that `async_` is false.
                // This blocking behavior resolves the race of `handle.abort()`
                // with the task actually getting cancelled or finishing.
                HostTaskState::CalleeRunning(handle) => {
                    handle.abort();
                    *state = HostTaskState::CalleeCancelling;
                    needs_block = true;
                }

                // Cancellation was already requested, so fail as the task can't
                // be cancelled twice.
                HostTaskState::CalleeCancelling | HostTaskState::CalleeDone { cancelled: true } => {
                    bail!(Trap::SubtaskCancelAfterTerminal);
                }
                HostTaskState::CalleeDone { cancelled: false } => {
                    // The callee is already done so there's no need to
                    // block further for an event.
                    *state = HostTaskState::CalleeDone { cancelled: true };
                    needs_block = false;
                }

                // These states should not be possible for a subtask that's
                // visible from the guest, so trap here.
                HostTaskState::CalleeStarted | HostTaskState::CalleeFinished(_) => {
                    bail_bug!("invalid states for host callee")
                }
            }
        } else {
            let guest_task = TableId::<GuestTask>::new(rep);
            let task = concurrent_state.get_mut(guest_task)?;

            // Cancellation was already requested, so fail as the task can't be
            // cancelled twice.
            if task.cancel_requested {
                bail!(Trap::SubtaskCancelAfterTerminal);
            }

            if !task.already_lowered_parameters() {
                store.cancel_guest_subtask_without_lowered_parameters(
                    self.runtime_instance(caller_instance),
                    guest_task,
                )?;
                return Ok(Status::StartCancelled as u32);
            } else if !task.returned_or_cancelled() {
                // Started, but not yet returned or cancelled; send the
                // `CANCELLED` event
                //
                // Note that this might overwrite an event that was set earlier
                // (e.g. `Event::None` if the task is yielding), but that's
                // okay -- this should supersede the previous state.
                task.cancel_requested = true;
                task.event = Some(Event::Cancelled);
                let runtime_instance = task.instance;
                for thread in task.threads.clone() {
                    let thread = QualifiedThreadId {
                        task: guest_task,
                        thread,
                    };
                    let concurrent_state = store.concurrent_state_mut()?;
                    let thread_mut = concurrent_state.get_mut(thread.thread)?;

                    let yield_ = |store: &mut StoreOpaque| {
                        // While we're yielding, temporarily set
                        // `do_not_suspend` to false on the subtask's instance
                        // since we'll be getting control back if it does
                        // suspend.
                        let state = store.instance_state(runtime_instance).concurrent_state();
                        let old_do_not_suspend = state.do_not_suspend;
                        state.do_not_suspend = false;

                        let caller = store.current_guest_thread()?;

                        // Temporarily add the waitable to the caller's
                        // `sync_call_set` to ensure that (1) it isn't already
                        // part of a different set and (2) it can't be added to
                        // a different set while we yield to the subtask.
                        let state = store.concurrent_state_mut()?;
                        let set = state.get_mut(caller.thread)?.sync_call_set;
                        waitable.join(state, Some(set))?;

                        store.suspend(SuspendReason::YieldingToSubtask { thread: caller })?;

                        let state = store.concurrent_state_mut()?;
                        waitable.join(state, None)?;

                        store
                            .instance_state(runtime_instance)
                            .concurrent_state()
                            .do_not_suspend = old_do_not_suspend;

                        Ok::<(), crate::Error>(())
                    };

                    match thread_mut.wake_on_cancel.take() {
                        WakeOnCancel::Waiting(set) => {
                            // The thread is in a cancellable wait, so wake it up:
                            let instance = concurrent_state.take_callback_waiter(set, thread)?;
                            let item = WorkItem::GuestCall {
                                instance: runtime_instance,
                                call: GuestCall {
                                    thread,
                                    kind: GuestCallKind::DeliverEvent {
                                        instance,
                                        set: Some(set),
                                    },
                                },
                            };
                            concurrent_state.set_switch_item(item)?;

                            yield_(store)?;

                            break;
                        }
                        WakeOnCancel::Yielding => {
                            if concurrent_state.promote_thread_work_item(thread)? {
                                yield_(store)?;
                                break;
                            } else if store
                                .instance_state(runtime_instance)
                                .concurrent_state()
                                .pending
                                .contains_key(&thread)
                            {
                                // The thread's work item was parked in
                                // `pending` because its instance can't be
                                // entered right now, so the cancellation
                                // can't be delivered yet. It'll be delivered
                                // via `GuestTask::event` once the instance
                                // becomes enterable and the work item runs.
                                // Until then the thread is still yielding, so
                                // restore `wake_on_cancel` and move on to the
                                // next thread.
                                store
                                    .concurrent_state_mut()?
                                    .get_mut(thread.thread)?
                                    .wake_on_cancel = WakeOnCancel::Yielding;
                            } else {
                                bail_bug!("thread with `WakeOnCancel::Yielding` not promotable");
                            }
                        }
                        WakeOnCancel::None => {}
                    }
                }

                // Guest tasks need to block if they have not yet returned or
                // cancelled, even as a result of the event delivery above.
                needs_block = !store
                    .concurrent_state_mut()?
                    .get_mut(guest_task)?
                    .returned_or_cancelled()
            } else {
                needs_block = false;
            }
        };

        // If we need to block waiting on the terminal status of this subtask
        // then return immediately in `async` mode, or otherwise wait for the
        // event to get signaled through the store.
        if needs_block {
            if async_ {
                return Ok(BLOCKED);
            }

            // Save and later restore `next_switch_item` during a sync cancel so
            // we don't try to switch to it while blocking.
            let old_next_switch_item = store.concurrent_state_mut()?.save_next_switch_item()?;

            // Wait for this waitable to get signaled with its terminal
            // status. Once that's done fall through to the shared code.
            store.wait_for_event(self.runtime_instance(caller_instance), waitable)?;

            let state = store.concurrent_state_mut()?;
            state.restore_next_switch_item(old_next_switch_item)?;

            // .. fall through to determine what event's in store for us.
        }

        let event = waitable.take_event(store.concurrent_state_mut()?)?;
        if let Some(Event::Subtask {
            status: status @ (Status::Returned | Status::ReturnCancelled),
        }) = event
        {
            Ok(status as u32)
        } else {
            bail!(Trap::SubtaskCancelAfterTerminal);
        }
    }
}

/// Trait representing component model ABI async intrinsics and fused adapter
/// helper functions.
///
/// SAFETY (callers): Most of the methods in this trait accept raw pointers,
/// which must be valid for at least the duration of the call (and possibly for
/// as long as the relevant guest task exists, in the case of `*mut VMFuncRef`
/// pointers used for async calls).
pub trait VMComponentAsyncStore {
    /// A helper function for fused adapter modules involving calls where the
    /// one of the caller or callee is async.
    ///
    /// This helper is not used when the caller and callee both use the sync
    /// ABI, only when at least one is async is this used.
    unsafe fn prepare_call(
        &mut self,
        instance: Instance,
        memory: *mut VMMemoryDefinition,
        start: NonNull<VMFuncRef>,
        return_: NonNull<VMFuncRef>,
        caller_instance: RuntimeComponentInstanceIndex,
        callee_instance: RuntimeComponentInstanceIndex,
        task_return_type: TypeTupleIndex,
        callee_async: bool,
        string_encoding: StringEncoding,
        result_count: u32,
        storage: *mut ValRaw,
        storage_len: usize,
    ) -> Result<()>;

    /// A helper function for fused adapter modules to start a call previously
    /// set up with `prepare_call`.
    unsafe fn start_call(
        &mut self,
        instance: Instance,
        callback: *mut VMFuncRef,
        post_return: *mut VMFuncRef,
        callee: NonNull<VMFuncRef>,
        param_count: u32,
        result_count: u32,
        flags: u32,
        storage: *mut MaybeUninit<ValRaw>,
        storage_len: usize,
    ) -> Result<()>;

    /// The `future.write` intrinsic.
    fn future_write(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        ty: TypeFutureTableIndex,
        options: OptionsIndex,
        future: u32,
        address: u32,
    ) -> Result<u32>;

    /// The `future.read` intrinsic.
    fn future_read(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        ty: TypeFutureTableIndex,
        options: OptionsIndex,
        future: u32,
        address: u32,
    ) -> Result<u32>;

    /// The `future.drop-writable` intrinsic.
    fn future_drop_writable(
        &mut self,
        instance: Instance,
        ty: TypeFutureTableIndex,
        writer: u32,
    ) -> Result<()>;

    /// The `stream.write` intrinsic.
    fn stream_write(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        ty: TypeStreamTableIndex,
        options: OptionsIndex,
        stream: u32,
        address: u32,
        count: u32,
    ) -> Result<u32>;

    /// The `stream.read` intrinsic.
    fn stream_read(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        ty: TypeStreamTableIndex,
        options: OptionsIndex,
        stream: u32,
        address: u32,
        count: u32,
    ) -> Result<u32>;

    /// The "fast-path" implementation of the `stream.write` intrinsic for
    /// "flat" (i.e. memcpy-able) payloads.
    fn flat_stream_write(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        ty: TypeStreamTableIndex,
        options: OptionsIndex,
        payload_size: u32,
        payload_align: u32,
        stream: u32,
        address: u32,
        count: u32,
    ) -> Result<u32>;

    /// The "fast-path" implementation of the `stream.read` intrinsic for "flat"
    /// (i.e. memcpy-able) payloads.
    fn flat_stream_read(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        ty: TypeStreamTableIndex,
        options: OptionsIndex,
        payload_size: u32,
        payload_align: u32,
        stream: u32,
        address: u32,
        count: u32,
    ) -> Result<u32>;

    /// The `stream.drop-writable` intrinsic.
    fn stream_drop_writable(
        &mut self,
        instance: Instance,
        ty: TypeStreamTableIndex,
        writer: u32,
    ) -> Result<()>;

    /// The `error-context.debug-message` intrinsic.
    fn error_context_debug_message(
        &mut self,
        instance: Instance,
        ty: TypeComponentLocalErrorContextTableIndex,
        options: OptionsIndex,
        err_ctx_handle: u32,
        debug_msg_address: u32,
    ) -> Result<()>;

    /// The `thread.new-indirect` intrinsic
    fn thread_new_indirect(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        func_ty_idx: ModuleInternedTypeIndex,
        start_func_table_idx: RuntimeTableIndex,
        start_func_idx: u32,
        context: i32,
    ) -> Result<u32>;
}

/// SAFETY: See trait docs.
impl<T: 'static> VMComponentAsyncStore for StoreInner<T> {
    unsafe fn prepare_call(
        &mut self,
        instance: Instance,
        memory: *mut VMMemoryDefinition,
        start: NonNull<VMFuncRef>,
        return_: NonNull<VMFuncRef>,
        caller_instance: RuntimeComponentInstanceIndex,
        callee_instance: RuntimeComponentInstanceIndex,
        task_return_type: TypeTupleIndex,
        callee_async: bool,
        string_encoding: StringEncoding,
        result_count_or_max_if_async: u32,
        storage: *mut ValRaw,
        storage_len: usize,
    ) -> Result<()> {
        // SAFETY: The `wasmtime_cranelift`-generated code that calls
        // this method will have ensured that `storage` is a valid
        // pointer containing at least `storage_len` items.
        let params = unsafe { core::slice::from_raw_parts(storage, storage_len) }.to_vec();

        unsafe {
            instance.prepare_call(
                StoreContextMut(self),
                start,
                return_,
                caller_instance,
                callee_instance,
                task_return_type,
                callee_async,
                memory,
                string_encoding,
                match result_count_or_max_if_async {
                    PREPARE_ASYNC_NO_RESULT => CallerInfo::Async {
                        params,
                        has_result: false,
                    },
                    PREPARE_ASYNC_WITH_RESULT => CallerInfo::Async {
                        params,
                        has_result: true,
                    },
                    result_count => CallerInfo::Sync {
                        params,
                        result_count,
                    },
                },
            )
        }
    }

    unsafe fn start_call(
        &mut self,
        instance: Instance,
        callback: *mut VMFuncRef,
        post_return: *mut VMFuncRef,
        callee: NonNull<VMFuncRef>,
        param_count: u32,
        result_count: u32,
        flags: u32,
        storage: *mut MaybeUninit<ValRaw>,
        storage_len: usize,
    ) -> Result<()> {
        unsafe {
            instance.start_call(
                StoreContextMut(self),
                callback,
                post_return,
                callee,
                param_count,
                result_count,
                flags,
                // SAFETY: The `wasmtime_cranelift`-generated code that calls
                // this method will have ensured that `storage` is a valid
                // pointer containing at least `storage_len` items.
                core::slice::from_raw_parts_mut(storage, storage_len),
            )
        }
    }

    fn future_write(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        ty: TypeFutureTableIndex,
        options: OptionsIndex,
        future: u32,
        address: u32,
    ) -> Result<u32> {
        instance
            .guest_write(
                StoreContextMut(self),
                caller,
                TransmitIndex::Future(ty),
                options,
                None,
                future,
                address,
                1,
            )
            .map(|result| result.encode())
    }

    fn future_read(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        ty: TypeFutureTableIndex,
        options: OptionsIndex,
        future: u32,
        address: u32,
    ) -> Result<u32> {
        instance
            .guest_read(
                StoreContextMut(self),
                caller,
                TransmitIndex::Future(ty),
                options,
                None,
                future,
                address,
                1,
            )
            .map(|result| result.encode())
    }

    fn stream_write(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        ty: TypeStreamTableIndex,
        options: OptionsIndex,
        stream: u32,
        address: u32,
        count: u32,
    ) -> Result<u32> {
        instance
            .guest_write(
                StoreContextMut(self),
                caller,
                TransmitIndex::Stream(ty),
                options,
                None,
                stream,
                address,
                count,
            )
            .map(|result| result.encode())
    }

    fn stream_read(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        ty: TypeStreamTableIndex,
        options: OptionsIndex,
        stream: u32,
        address: u32,
        count: u32,
    ) -> Result<u32> {
        instance
            .guest_read(
                StoreContextMut(self),
                caller,
                TransmitIndex::Stream(ty),
                options,
                None,
                stream,
                address,
                count,
            )
            .map(|result| result.encode())
    }

    fn future_drop_writable(
        &mut self,
        instance: Instance,
        ty: TypeFutureTableIndex,
        writer: u32,
    ) -> Result<()> {
        instance.guest_drop_writable(self, TransmitIndex::Future(ty), writer)
    }

    fn flat_stream_write(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        ty: TypeStreamTableIndex,
        options: OptionsIndex,
        payload_size: u32,
        payload_align: u32,
        stream: u32,
        address: u32,
        count: u32,
    ) -> Result<u32> {
        instance
            .guest_write(
                StoreContextMut(self),
                caller,
                TransmitIndex::Stream(ty),
                options,
                Some(FlatAbi {
                    size: payload_size,
                    align: payload_align,
                }),
                stream,
                address,
                count,
            )
            .map(|result| result.encode())
    }

    fn flat_stream_read(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        ty: TypeStreamTableIndex,
        options: OptionsIndex,
        payload_size: u32,
        payload_align: u32,
        stream: u32,
        address: u32,
        count: u32,
    ) -> Result<u32> {
        instance
            .guest_read(
                StoreContextMut(self),
                caller,
                TransmitIndex::Stream(ty),
                options,
                Some(FlatAbi {
                    size: payload_size,
                    align: payload_align,
                }),
                stream,
                address,
                count,
            )
            .map(|result| result.encode())
    }

    fn stream_drop_writable(
        &mut self,
        instance: Instance,
        ty: TypeStreamTableIndex,
        writer: u32,
    ) -> Result<()> {
        instance.guest_drop_writable(self, TransmitIndex::Stream(ty), writer)
    }

    fn error_context_debug_message(
        &mut self,
        instance: Instance,
        ty: TypeComponentLocalErrorContextTableIndex,
        options: OptionsIndex,
        err_ctx_handle: u32,
        debug_msg_address: u32,
    ) -> Result<()> {
        instance.error_context_debug_message(
            StoreContextMut(self),
            ty,
            options,
            err_ctx_handle,
            debug_msg_address,
        )
    }

    fn thread_new_indirect(
        &mut self,
        instance: Instance,
        caller: RuntimeComponentInstanceIndex,
        func_ty_idx: ModuleInternedTypeIndex,
        start_func_table_idx: RuntimeTableIndex,
        start_func_idx: u32,
        context: i32,
    ) -> Result<u32> {
        instance.thread_new_indirect(
            StoreContextMut(self),
            caller,
            func_ty_idx,
            start_func_table_idx,
            start_func_idx,
            context,
        )
    }
}

type HostTaskFuture = Pin<Box<dyn Future<Output = Result<()>> + Send + 'static>>;

/// Runs the given future with the current thread set to `task` each time it is
/// polled.
async fn run_with_host_task_set<F>(task: TableId<HostTask>, future: F) -> Result<F::Output>
where
    F: Future,
{
    let mut future = pin!(future);
    future::poll_fn(|cx| {
        let old_thread = match tls::get(|store| store.set_thread(task)) {
            Ok(thread) => thread,
            Err(error) => return Poll::Ready(Err(error)),
        };
        let result = future.as_mut().poll(cx);
        match tls::get(|store| store.set_thread(old_thread)) {
            Ok(_) => result.map(Ok),
            Err(error) => Poll::Ready(Err(error)),
        }
    })
    .await
}

/// Represents the state of a pending host task.
///
/// This is used to represent tasks when the guest calls into the host.
pub(crate) struct HostTask {
    common: WaitableCommon,

    /// State of borrows/etc the host needs to track. Used when the guest passes
    /// borrows to the host, for example.
    call_context: CallContext,

    state: HostTaskState,

    group: TaskGroupId,
}

enum HostTaskState {
    /// A host task has been created and it's considered "started".
    ///
    /// The host task has yet to enter `first_poll` or `poll_and_block` which
    /// is where this will get updated further.
    CalleeStarted,

    /// State used for tasks in `first_poll` meaning that the guest did an async
    /// lower of a host async function which is blocked. The specified handle is
    /// linked to the future in the main `FuturesUnordered` of a store which is
    /// used to cancel it if the guest requests cancellation.
    CalleeRunning(JoinHandle),

    /// Cancellation was requested, but the completion worker has not finished
    /// lowering the result and publishing the terminal event. The task must
    /// remain in the table until that worker completes.
    CalleeCancelling,

    /// Terminal state used for tasks in `poll_and_block` to store the result of
    /// their computation. Note that this state is not used for tasks in
    /// `first_poll`.
    CalleeFinished(LiftedResult),

    /// Terminal state for host tasks meaning that the task was cancelled or the
    /// result was taken.
    CalleeDone { cancelled: bool },
}

impl HostTask {
    fn new(
        concurrent_state: &mut ConcurrentState,
        state: HostTaskState,
        caller: QualifiedThreadId,
    ) -> Result<Self> {
        let group = concurrent_state.get_mut(caller.task)?.group;
        concurrent_state.increment_group_ref_count(group)?;

        Ok(Self {
            common: WaitableCommon::default(),
            call_context: CallContext::default(),
            state,
            group,
        })
    }
}

impl TableDebug for HostTask {
    fn type_name() -> &'static str {
        "HostTask"
    }
}

type CallbackFn = Box<dyn Fn(&mut dyn VMStore, Event, u32) -> Result<u32> + Send + Sync + 'static>;

/// Represents the caller of a given guest task.
enum Caller {
    /// The host called the guest task.
    Host {
        /// If present, may be used to deliver the result.
        tx: Option<oneshot::Sender<LiftedResult>>,
        /// The host task which called into the guest, or `None` for a call from
        /// the top-level host. The task may belong to an entirely unrelated
        /// top-level component instance than the one the host called into.
        caller: Option<TableId<HostTask>>,
    },
    /// Another guest thread called the guest task
    Guest {
        /// The id of the caller
        thread: QualifiedThreadId,
    },
}

/// Represents a closure and related canonical ABI parameters required to
/// validate a `task.return` call at runtime and lift the result.
struct LiftResult {
    lift: RawLift,
    ty: TypeTupleIndex,
    memory: Option<SendSyncPtr<VMMemoryDefinition>>,
    string_encoding: StringEncoding,
}

/// The table ID for a guest thread, qualified by the task to which it belongs.
///
/// This exists to minimize table lookups and the necessity to pass stores around mutably
/// for the common case of identifying the task to which a thread belongs.
#[derive(Copy, Clone, Ord, PartialOrd, Eq, PartialEq)]
pub(crate) struct QualifiedThreadId {
    task: TableId<GuestTask>,
    thread: TableId<GuestThread>,
}

impl QualifiedThreadId {
    fn qualify(
        state: &mut ConcurrentState,
        thread: TableId<GuestThread>,
    ) -> Result<QualifiedThreadId> {
        Ok(QualifiedThreadId {
            task: state.get_mut(thread)?.parent_task,
            thread,
        })
    }
}

impl fmt::Debug for QualifiedThreadId {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.debug_tuple("QualifiedThreadId")
            .field(&self.task.rep())
            .field(&self.thread.rep())
            .finish()
    }
}

enum GuestThreadState {
    NotStartedImplicit,
    NotStartedExplicit(
        Box<dyn FnOnce(&mut dyn VMStore, QualifiedThreadId) -> Result<()> + Send + Sync>,
    ),
    Running,
    Suspended(StoreFiber<'static>),
    Ready {
        fiber: StoreFiber<'static>,
    },
    Completed,
}

impl GuestThreadState {
    fn check_no_fiber(&self) -> Result<()> {
        match self {
            Self::Suspended(_) | Self::Ready { .. } => bail_bug!("thread already owns a fiber"),
            Self::NotStartedImplicit
            | Self::NotStartedExplicit(_)
            | Self::Running
            | Self::Completed => Ok(()),
        }
    }
}

impl fmt::Debug for GuestThreadState {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            Self::NotStartedImplicit => f.debug_tuple("NotStartedImplicit").finish(),
            Self::NotStartedExplicit(_) => f.debug_tuple("NotStartedExplicit").finish(),
            Self::Running => f.debug_tuple("Running").finish(),
            Self::Suspended(_) => f.debug_tuple("Suspended").finish(),
            Self::Ready { .. } => f.debug_struct("Ready").finish(),
            Self::Completed => f.debug_tuple("Completed").finish(),
        }
    }
}

#[derive(Copy, Clone, PartialEq, Eq, Debug)]
enum WakeOnCancel {
    None,
    Waiting(TableId<WaitableSet>),
    Yielding,
}

impl WakeOnCancel {
    fn is_none(self) -> bool {
        matches!(self, WakeOnCancel::None)
    }

    fn replace(&mut self, other: WakeOnCancel) -> Self {
        let old = *self;
        *self = other;
        old
    }

    fn take(&mut self) -> Self {
        self.replace(WakeOnCancel::None)
    }
}

pub struct GuestThread {
    /// Context-local state used to implement the `context.{get,set}`
    /// intrinsics.
    context: [u32; NUM_COMPONENT_CONTEXT_SLOTS],
    /// The owning guest task.
    parent_task: TableId<GuestTask>,
    /// If non-`None`, indicates that the thread is currently either waiting a
    /// waitable set or yielding but may be cancelled and woken immediately.
    wake_on_cancel: WakeOnCancel,
    /// The execution state of this guest thread
    state: GuestThreadState,
    /// The index of this thread in the component instance's handle table.
    /// This must always be `Some` after initialization.
    instance_rep: Option<u32>,
    /// Scratch waitable set used to watch subtasks during synchronous calls.
    sync_call_set: TableId<WaitableSet>,
    /// The old value of `do_not_suspend` prior to the sync-typed task for which
    /// this thread was created, if relevant.
    old_do_not_suspend: Option<bool>,
}

impl GuestThread {
    /// Retrieve the `GuestThread` corresponding to the specified guest-visible
    /// handle.
    fn from_instance(
        state: Pin<&mut ComponentInstance>,
        caller_instance: RuntimeComponentInstanceIndex,
        guest_thread: u32,
    ) -> Result<TableId<Self>> {
        let rep = state.instance_states().0[caller_instance]
            .thread_handle_table()
            .guest_thread_rep(guest_thread)?;
        Ok(TableId::new(rep))
    }

    fn new_implicit(state: &mut ConcurrentState, parent_task: TableId<GuestTask>) -> Result<Self> {
        let sync_call_set = state.push(WaitableSet {
            is_sync_call_set: true,
            ..WaitableSet::default()
        })?;
        Ok(Self {
            context: [0; NUM_COMPONENT_CONTEXT_SLOTS],
            parent_task,
            wake_on_cancel: WakeOnCancel::None,
            state: GuestThreadState::NotStartedImplicit,
            instance_rep: None,
            sync_call_set,
            old_do_not_suspend: None,
        })
    }

    fn new_explicit(
        state: &mut ConcurrentState,
        parent_task: TableId<GuestTask>,
        start_func: Box<
            dyn FnOnce(&mut dyn VMStore, QualifiedThreadId) -> Result<()> + Send + Sync,
        >,
    ) -> Result<Self> {
        state.increment_task_ref_count(parent_task)?;
        let sync_call_set = state.push(WaitableSet {
            is_sync_call_set: true,
            ..WaitableSet::default()
        })?;
        Ok(Self {
            context: [0; NUM_COMPONENT_CONTEXT_SLOTS],
            parent_task,
            wake_on_cancel: WakeOnCancel::None,
            state: GuestThreadState::NotStartedExplicit(start_func),
            instance_rep: None,
            sync_call_set,
            old_do_not_suspend: None,
        })
    }
}

impl TableDebug for GuestThread {
    fn type_name() -> &'static str {
        "GuestThread"
    }
}

enum SyncResult {
    NotProduced,
    Produced(Option<ValRaw>),
    Taken,
}

impl SyncResult {
    fn take(&mut self) -> Result<Option<Option<ValRaw>>> {
        Ok(match mem::replace(self, SyncResult::Taken) {
            SyncResult::NotProduced => None,
            SyncResult::Produced(val) => Some(val),
            SyncResult::Taken => {
                bail_bug!("attempted to take a synchronous result that was already taken")
            }
        })
    }
}

/// Represents a pending guest task.
pub(crate) struct GuestTask {
    /// See `WaitableCommon`
    common: WaitableCommon,
    /// Closure to lower the parameters passed to this task.
    lower_params: Option<RawLower>,
    /// See `LiftResult`
    lift_result: Option<LiftResult>,
    /// A place to stash the type-erased lifted result if it can't be delivered
    /// immediately.
    result: Option<LiftedResult>,
    /// Closure to call the callback function for an async-lifted export, if
    /// provided.
    callback: Option<CallbackFn>,
    /// See `Caller`
    caller: Caller,
    /// Borrow state for this task.
    ///
    /// Keeps track of `borrow<T>` received to this task to ensure that
    /// everything is dropped by the time it exits.
    call_context: CallContext,
    /// A place to stash the lowered result for a sync-to-async call until it
    /// can be returned to the caller.
    sync_result: SyncResult,
    /// Whether or not the task has been cancelled (i.e. whether the
    /// cancellation request has been delivered to the task, and thus whether
    /// the task is permitted to call `task.cancel`).
    cancel_request_delivered: bool,
    /// Whether or not the caller has requested cancellation of this task via
    /// `subtask.cancel` while it was still unresolved. A subsequent
    /// `subtask.cancel` of the same task traps. This matches the spec's
    /// `Subtask.cancellation_requested`.
    cancel_requested: bool,
    /// Whether or not we've sent a `Status::Starting` event to any current or
    /// future waiters for this waitable.
    starting_sent: bool,
    /// The runtime instance to which the exported function for this guest task
    /// belongs.
    ///
    /// Note that the task may do a sync->sync call via a fused adapter which
    /// results in that task executing code in a different instance, and it may
    /// call host functions and intrinsics from that other instance.
    instance: RuntimeInstance,
    /// If present, a pending `Event::Cancelled` to be delivered to this task.
    event: Option<Event>,
    /// Whether or not the task has exited.
    exited: bool,
    /// Threads belonging to this task
    threads: HashSet<TableId<GuestThread>>,
    /// Indicates whether this task was created for a call to an async-typed
    /// export.
    async_typed: bool,
    /// Indicates whether this task was created for a call to an async-lifted
    /// export.
    async_lifted: bool,
    /// Whether `ConcurrentState::interesting_tasks` has been decremented for
    /// this task yet.
    decremented_interesting_task_count: bool,
    /// The `TaskGroup` to which this task belongs.
    group: TaskGroupId,
    /// The number of outstanding references to this task.
    ///
    /// When this goes to zero, the task will be removed from
    /// `ConcurrentState::table`.
    ref_count: usize,
}

impl GuestTask {
    fn already_lowered_parameters(&self) -> bool {
        // We reset `lower_params` after we lower the parameters
        self.lower_params.is_none()
    }

    fn returned_or_cancelled(&self) -> bool {
        // We reset `lift_result` after we return or exit
        self.lift_result.is_none()
    }

    fn new(
        state: &mut ConcurrentState,
        lower_params: RawLower,
        lift_result: LiftResult,
        caller: Caller,
        callback: Option<CallbackFn>,
        instance: RuntimeInstance,
        async_typed: bool,
        async_lifted: bool,
    ) -> Result<QualifiedThreadId> {
        let group = match caller {
            Caller::Guest { thread } => {
                let group = state.get_mut(thread.task)?.group;
                state.increment_group_ref_count(group)?;
                group
            }
            Caller::Host { .. } => state.make_task_group()?,
        };

        let task = state.push(Self {
            common: WaitableCommon::default(),
            lower_params: Some(lower_params),
            lift_result: Some(lift_result),
            result: None,
            callback,
            caller,
            call_context: CallContext::default(),
            sync_result: SyncResult::NotProduced,
            cancel_request_delivered: false,
            cancel_requested: false,
            starting_sent: false,
            instance,
            event: None,
            exited: false,
            threads: HashSet::new(),
            async_typed,
            async_lifted,
            decremented_interesting_task_count: false,
            group,
            // One for the caller and one for the implicit guest thread we're
            // about to create:
            ref_count: 2,
        })?;
        let new_thread = GuestThread::new_implicit(state, task)?;
        let thread = state.push(new_thread)?;
        state.get_mut(task)?.threads.insert(thread);
        state.interesting_tasks += 1;
        let thread = QualifiedThreadId { task, thread };
        log::trace!("new implicit thread {thread:?} for instance {instance:?}");
        Ok(thread)
    }
}

impl TableDebug for GuestTask {
    fn type_name() -> &'static str {
        "GuestTask"
    }
}

/// Represents state common to all kinds of waitables.
#[derive(Default)]
struct WaitableCommon {
    /// The currently pending event for this waitable, if any.
    event: Option<Event>,
    /// The set to which this waitable belongs, if any.
    set: Option<TableId<WaitableSet>>,
    /// The handle with which the guest refers to this waitable, if any.
    handle: Option<u32>,
}

/// Represents a Component Model Async `waitable`.
#[derive(Copy, Clone, Ord, PartialOrd, Eq, PartialEq)]
enum Waitable {
    /// A host task
    Host(TableId<HostTask>),
    /// A guest task
    Guest(TableId<GuestTask>),
    /// The read or write end of a stream or future
    Transmit(TableId<TransmitHandle>),
}

impl Waitable {
    /// Retrieve the `Waitable` corresponding to the specified guest-visible
    /// handle.
    fn from_instance(
        state: Pin<&mut ComponentInstance>,
        caller_instance: RuntimeComponentInstanceIndex,
        waitable: u32,
    ) -> Result<Self> {
        use crate::runtime::vm::component::Waitable;

        let (waitable, kind) = state.instance_states().0[caller_instance]
            .handle_table()
            .waitable_rep(waitable)?;

        Ok(match kind {
            Waitable::Subtask { is_host: true } => Self::Host(TableId::new(waitable)),
            Waitable::Subtask { is_host: false } => Self::Guest(TableId::new(waitable)),
            Waitable::Stream | Waitable::Future => Self::Transmit(TableId::new(waitable)),
        })
    }

    /// Retrieve the host-visible identifier for this `Waitable`.
    fn rep(&self) -> u32 {
        match self {
            Self::Host(id) => id.rep(),
            Self::Guest(id) => id.rep(),
            Self::Transmit(id) => id.rep(),
        }
    }

    /// Move this `Waitable` to the specified set (when `set` is `Some(_)`) or
    /// remove it from any set it may currently belong to (when `set` is
    /// `None`).
    fn join(&self, state: &mut ConcurrentState, set: Option<TableId<WaitableSet>>) -> Result<()> {
        log::trace!("waitable {self:?} join set {set:?}");

        let old = mem::replace(&mut self.common(state)?.set, set);

        if let Some(old) = old {
            match *self {
                Waitable::Host(id) => state.remove_child(id, old),
                Waitable::Guest(id) => state.remove_child(id, old),
                Waitable::Transmit(id) => state.remove_child(id, old),
            }?;

            state.get_mut(old)?.ready.remove(self);
        }

        if let Some(set) = set {
            match *self {
                Waitable::Host(id) => state.add_child(id, set),
                Waitable::Guest(id) => state.add_child(id, set),
                Waitable::Transmit(id) => state.add_child(id, set),
            }?;

            if self.common(state)?.event.is_some() {
                self.mark_ready(state)?;
            }
        }

        Ok(())
    }

    /// Retrieve mutable access to the `WaitableCommon` for this `Waitable`.
    fn common<'a>(&self, state: &'a mut ConcurrentState) -> Result<&'a mut WaitableCommon> {
        Ok(match self {
            Self::Host(id) => &mut state.get_mut(*id)?.common,
            Self::Guest(id) => &mut state.get_mut(*id)?.common,
            Self::Transmit(id) => &mut state.get_mut(*id)?.common,
        })
    }

    /// Trap if this waitable is currently a member of a waitable set.
    ///
    /// A synchronous stream/future/subtask operation may end up blocking on
    /// this waitable, so it is not allowed to run while the waitable is also
    /// being watched by a waitable set.
    fn trap_if_in_waitable_set(&self, state: &mut ConcurrentState) -> Result<()> {
        if self.common(state)?.set.is_some() {
            bail!(Trap::WaitableSyncAndAsync);
        }
        Ok(())
    }

    /// Trap if a thread is currently synchronously waiting on this waitable.
    ///
    /// This corresponds to the spec's `trap_if(w.has_sync_waiter)` and is
    /// used to prevent other threads from stealing or otherwise interfering
    /// with the event the sync waiter is expecting.
    fn trap_if_has_sync_waiter(&self, state: &mut ConcurrentState) -> Result<()> {
        if let Some(set) = self.common(state)?.set
            && state.get_mut(set)?.is_sync_call_set
        {
            bail!(Trap::WaitableSyncAndAsync);
        }
        Ok(())
    }

    /// Set or clear the pending event for this waitable and either deliver it
    /// to the first waiter, if any, or mark it as ready to be delivered to the
    /// next waiter that arrives.
    fn set_event(&self, state: &mut ConcurrentState, event: Option<Event>) -> Result<()> {
        log::trace!("set event for {self:?}: {event:?}");
        self.common(state)?.event = event;
        self.mark_ready(state)
    }

    /// Take the pending event from this waitable, leaving `None` in its place.
    fn take_event(&self, state: &mut ConcurrentState) -> Result<Option<Event>> {
        let common = self.common(state)?;
        let event = common.event.take();
        if let Some(set) = self.common(state)?.set {
            state.get_mut(set)?.ready.remove(self);
        }

        Ok(event)
    }

    /// Deliver the current event for this waitable to the first waiter, if any,
    /// or else mark it as ready to be delivered to the next waiter that
    /// arrives.
    fn mark_ready(&self, state: &mut ConcurrentState) -> Result<()> {
        if let Some(set) = self.common(state)?.set {
            state.get_mut(set)?.ready.insert(*self);
            state.wake_waiter(set)?;
        }
        Ok(())
    }

    /// Remove this waitable from the store's rep table (or decrement the ref
    /// count if it's a guest task).
    fn delete_from(&self, state: &mut ConcurrentState) -> Result<()> {
        match self {
            Self::Host(task) => state.delete_host_task(*task),
            Self::Guest(task) => state.decrement_task_ref_count(*task),
            Self::Transmit(task) => Ok(state.delete(*task).map(drop)?),
        }
    }
}

impl fmt::Debug for Waitable {
    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
        match self {
            Self::Host(id) => write!(f, "{id:?}"),
            Self::Guest(id) => write!(f, "{id:?}"),
            Self::Transmit(id) => write!(f, "{id:?}"),
        }
    }
}

/// Represents a Component Model Async `waitable-set`.
#[derive(Default)]
struct WaitableSet {
    /// Which waitables in this set have pending events, if any.
    ready: BTreeSet<Waitable>,
    /// Which guest threads are currently waiting on this set, if any.
    waiting: BTreeMap<QualifiedThreadId, WaitMode>,
    /// How many guest threads are waiting on this set, mirroring the spec's
    /// `WaitableSet.num_waiting`.
    num_waiting: usize,
    /// Whether this set is a synthetic, internal one meant for handling
    /// synchronous calls.
    is_sync_call_set: bool,
}

impl WaitableSet {
    /// Records that a thread counted in `num_waiting` is no longer waiting.
    fn stop_waiting(&mut self) -> Result<()> {
        self.num_waiting = match self.num_waiting.checked_sub(1) {
            Some(n) => n,
            None => bail_bug!("waiter not accounted for in waitable set"),
        };
        Ok(())
    }
}

impl TableDebug for WaitableSet {
    fn type_name() -> &'static str {
        "WaitableSet"
    }
}

/// Type-erased closure to lower the parameters for a guest task.
type RawLower =
    Box<dyn FnOnce(&mut dyn VMStore, &mut [MaybeUninit<ValRaw>]) -> Result<()> + Send + Sync>;

/// Type-erased closure to lift the result for a guest task.
type RawLift = Box<
    dyn FnOnce(&mut dyn VMStore, &[ValRaw]) -> Result<Box<dyn Any + Send + Sync>> + Send + Sync,
>;

/// Type erased result of a guest task which may be downcast to the expected
/// type by a host caller (or simply ignored in the case of a guest caller; see
/// `DummyResult`).
type LiftedResult = Box<dyn Any + Send + Sync>;

/// Used to return a result from a `LiftFn` when the actual result has already
/// been lowered to a guest task's stack and linear memory.
struct DummyResult;

/// Represents the Component Model Async state of a (sub-)component instance.
#[derive(Default)]
pub struct ConcurrentInstanceState {
    /// Whether backpressure is set for this instance (enabled if >0)
    backpressure: u16,
    /// Whether this instance can be entered
    do_not_enter: bool,
    /// Whether this instance may suspend (i.e. whether this instance is
    /// currently running a sync-typed function).
    do_not_suspend: bool,
    /// Pending calls for this instance which require `Self::backpressure` to be
    /// zero and/or `Self::do_not_enter` to be false before they can proceed.
    pending: BTreeMap<QualifiedThreadId, GuestCallKind>,
}

impl ConcurrentInstanceState {
    pub fn pending_is_empty(&self) -> bool {
        self.pending.is_empty()
    }
}

#[derive(Debug, Copy, Clone)]
pub(crate) enum CurrentThread {
    /// The currently running thread is a guest, identified here with its
    /// task/thread id combo.
    Guest(QualifiedThreadId),
    /// The currently running thread is a host task.
    Host(TableId<HostTask>),
    /// The currently running thread is a host call whose task has not yet been
    /// materialized. The contained ID identifies its guest caller.
    DeferredHost(QualifiedThreadId),
    /// There is no currently running thread because we are in the main event
    /// loop or concurrency is disabled.
    None,
}

impl CurrentThread {
    fn guest(&self) -> Option<&QualifiedThreadId> {
        match self {
            Self::Guest(id) => Some(id),
            _ => None,
        }
    }

    fn guest_task(&self) -> Option<TableId<GuestTask>> {
        match self {
            Self::Guest(id) => Some(id.task),
            _ => None,
        }
    }

    fn is_none(&self) -> bool {
        matches!(self, Self::None)
    }
}

impl From<QualifiedThreadId> for CurrentThread {
    fn from(id: QualifiedThreadId) -> Self {
        Self::Guest(id)
    }
}

impl From<TableId<HostTask>> for CurrentThread {
    fn from(id: TableId<HostTask>) -> Self {
        Self::Host(id)
    }
}

enum Priority {
    Switch,
    High,
    Low,
}

/// Represents the Component Model Async state of a store.
pub struct ConcurrentState {
    /// The currently running thread, if any.
    ///
    /// Note that we lazily materialize threads on-demand and this field is not
    /// necessarily up-to-date. The `StoreOpaque::current_thread` method should
    /// be preferred over directly accessing this field.
    unforced_current_thread: CurrentThread,

    /// Borrow state for the deferred host call, if any.
    ///
    /// This is `Some` if and only if [`Self::unforced_current_thread`] is
    /// [`CurrentThread::DeferredHost`]. Materializing the host task moves this
    /// context into that task.
    deferred_host_call_context: Option<CallContext>,

    /// The set of pending host and background tasks, if any.
    ///
    /// See `ComponentInstance::poll_until` for where we temporarily take this
    /// out, poll it, then put it back to avoid any mutable aliasing hazards.
    futures: AlwaysMut<Option<FuturesUnordered<HostTaskFuture>>>,
    /// The table of waitables, waitable sets, etc.
    table: AlwaysMut<ResourceTable>,
    /// The next item to switch to if any.
    ///
    /// This takes precedence over items in the `high_priority` queue below and
    /// should be used in cases such as subtask calls and thread resume/promote
    /// operations where we must switch to a specific thread at the next turn of
    /// the event loop regardless of what happens to be present in the
    /// `high_priority` queue.
    switch_item: Option<WorkItem>,
    /// The item to set `switch_item` to when the current thread suspends.
    ///
    /// This is used when ever an async-lowered import or `subtask.cancel` is
    /// called in order to track the thread to switch back to once the current
    /// subtask suspends, if any.
    next_switch_item: Option<WorkItem>,
    /// The "high priority" work queue for this store's event loop.
    high_priority: VecDeque<WorkItem>,
    /// The "low priority" work queue for this store's event loop.
    low_priority: VecDeque<WorkItem>,
    /// A place to stash the reason a fiber is suspending so that the code which
    /// resumed it will know under what conditions the fiber should be resumed
    /// again.
    suspend_reason: Option<SuspendReason>,
    /// A cached fiber which is waiting for work to do.
    ///
    /// This helps us avoid creating a new fiber for each `GuestCall` work item.
    worker: Option<StoreFiber<'static>>,
    /// A place to stash the work item for which we're resuming a worker fiber.
    worker_item: Option<WorkerItem>,

    /// Reference counts for all component error contexts
    ///
    /// NOTE: it is possible the global ref count to be *greater* than the sum of
    /// (sub)component ref counts as tracked by `error_context_tables`, for
    /// example when the host holds one or more references to error contexts.
    ///
    /// The key of this primary map is often referred to as the "rep" (i.e. host-side
    /// component-wide representation) of the index into concurrent state for a given
    /// stored `ErrorContext`.
    ///
    /// Stated another way, `TypeComponentGlobalErrorContextTableIndex` is essentially the same
    /// as a `TableId<ErrorContextState>`.
    global_error_context_ref_counts:
        BTreeMap<TypeComponentGlobalErrorContextTableIndex, GlobalErrorContextRefCount>,

    /// The number of "interesting tasks" currently executing in the store.
    ///
    /// This tracks the concept of a component instance lifetime as defined in
    /// https://github.com/WebAssembly/component-model/pull/643. Specifically
    /// all tasks currently increment this counter which then gets decremented
    /// when they exit. In the future some tasks might not increment this
    /// counter, but for now all do.
    ///
    /// This is used to implement `Accessor::poll_no_interesting_tasks` to
    /// inform the embedder when all tasks have completed. This is then
    /// used in wasmtime-wasi-http, for example, to know when an instance is
    /// idle.
    interesting_tasks: usize,

    /// Single waker to notify when `interesting_tasks` reaches 0.
    ///
    /// Used in the implementation of `Accessor::poll_no_interesting_tasks`.
    interesting_tasks_empty_waker: Option<Waker>,

    /// Single waker to notify when a component instance goes from
    /// not-concurrently-callable to concurrently-callable.
    ///
    /// Used in the implementation of `Accessor::poll_ready_for_concurrent_call`.
    ready_for_concurrent_call_waker: Option<Waker>,

    /// Whether the `StoreContextMut::poll_until` event loop is running.
    event_loop_running: bool,

    /// Stack of switch items to push to and pop from when entering and exiting
    /// sync-to-sync calls.
    saved_next_switch_items: Vec<Option<WorkItem>>,

    /// See [TaskGroupHook].
    #[cfg(feature = "task-group-hook")]
    task_group_hook: Option<Box<dyn TaskGroupHook>>,
}

impl Default for ConcurrentState {
    fn default() -> Self {
        Self {
            unforced_current_thread: CurrentThread::None,
            deferred_host_call_context: None,
            table: AlwaysMut::new(ResourceTable::new()),
            futures: AlwaysMut::new(Some(FuturesUnordered::new())),
            switch_item: None,
            next_switch_item: None,
            high_priority: VecDeque::new(),
            low_priority: VecDeque::new(),
            suspend_reason: None,
            worker: None,
            worker_item: None,
            global_error_context_ref_counts: BTreeMap::new(),
            interesting_tasks: 0,
            interesting_tasks_empty_waker: None,
            ready_for_concurrent_call_waker: None,
            event_loop_running: false,
            saved_next_switch_items: Vec::new(),
            #[cfg(feature = "task-group-hook")]
            task_group_hook: None,
        }
    }
}

impl ConcurrentState {
    /// Take ownership of any fibers and futures owned by this object.
    ///
    /// This should be used when disposing of the `Store` containing this object
    /// in order to gracefully resolve any and all fibers using
    /// `StoreFiber::dispose`.  This is necessary to avoid possible
    /// use-after-free bugs due to fibers which may still have access to the
    /// `Store`.
    ///
    /// Additionally, the futures collected with this function should be dropped
    /// within a `tls::set` call, which will ensure than any futures closing
    /// over an `&Accessor` will have access to the store when dropped, allowing
    /// e.g. `WithAccessor[AndValue]` instances to be disposed of without
    /// panicking.
    ///
    /// Note that this will leave the object in an inconsistent and unusable
    /// state, so it should only be used just prior to dropping it.
    pub(crate) fn take_fibers_and_futures(
        &mut self,
        fibers: &mut Vec<StoreFiber<'static>>,
        futures: &mut Vec<FuturesUnordered<HostTaskFuture>>,
    ) {
        let ConcurrentState {
            table,
            worker,
            switch_item,
            next_switch_item,
            high_priority,
            low_priority,
            saved_next_switch_items,
            futures: pending_futures,
            worker_item: _,
            unforced_current_thread: _,
            deferred_host_call_context: _,
            suspend_reason: _,
            global_error_context_ref_counts: _,
            interesting_tasks: _,
            interesting_tasks_empty_waker: _,
            ready_for_concurrent_call_waker: _,
            event_loop_running: _,
            #[cfg(feature = "task-group-hook")]
                task_group_hook: _,
        } = self;

        let mut items = Vec::new();
        for (_, entry) in table.get_mut().iter_mut() {
            if let Some(set) = entry.downcast_mut::<WaitableSet>() {
                for mode in mem::take(&mut set.waiting).into_values() {
                    match mode {
                        WaitMode::Fiber(fiber) => {
                            fibers.push(fiber);
                        }
                        WaitMode::Callback(_) => {}
                    }
                }
            } else if let Some(thread) = entry.downcast_mut::<GuestThread>() {
                match mem::replace(&mut thread.state, GuestThreadState::Completed) {
                    GuestThreadState::Suspended(fiber) | GuestThreadState::Ready { fiber } => {
                        fibers.push(fiber);
                    }
                    GuestThreadState::NotStartedImplicit
                    | GuestThreadState::NotStartedExplicit(_)
                    | GuestThreadState::Running
                    | GuestThreadState::Completed => {}
                }
            } else if let Some(item) = entry.downcast_mut::<Option<WorkItem>>() {
                if let Some(item) = item.take() {
                    items.push(item);
                }
            }
        }

        if let Some(fiber) = worker.take() {
            fibers.push(fiber);
        }

        let mut handle_item = |item| match item {
            WorkItem::ResumeFiber { fiber, .. } => {
                fibers.push(fiber);
            }
            WorkItem::PushFuture(future) => {
                pending_futures
                    .get_mut()
                    .as_mut()
                    .unwrap()
                    .push(future.into_inner());
            }
            WorkItem::ResumeThread { .. }
            | WorkItem::GuestCall { .. }
            | WorkItem::WorkerFunction(_) => {}
        };

        for item in items {
            handle_item(item);
        }
        if let Some(item) = switch_item.take() {
            handle_item(item);
        }
        if let Some(item) = next_switch_item.take() {
            handle_item(item);
        }
        for item in mem::take(high_priority) {
            handle_item(item);
        }
        for item in mem::take(low_priority) {
            handle_item(item);
        }
        for item in mem::take(saved_next_switch_items)
            .into_iter()
            .filter_map(|v| v)
        {
            handle_item(item);
        }

        if let Some(them) = pending_futures.get_mut().take() {
            futures.push(them);
        }
    }

    #[cfg(feature = "gc")]
    pub(crate) fn trace_fiber_roots(
        &mut self,
        modules: &ModuleRegistry,
        unwind: &dyn Unwind,
        gc_roots_list: &mut GcRootsList,
    ) {
        let ConcurrentState {
            table,
            worker,
            switch_item,
            next_switch_item,
            high_priority,
            low_priority,
            saved_next_switch_items,

            // TODO(cm-gc): This field contains `ValRaw`s, but they are never GC
            // references because the component model doesn't support GC yet. We
            // will need to trace these somehow when it does.
            futures: _,

            // These fields do not contain GC references.
            worker_item: _,
            unforced_current_thread: _,
            deferred_host_call_context: _,
            suspend_reason: _,
            global_error_context_ref_counts: _,
            interesting_tasks: _,
            interesting_tasks_empty_waker: _,
            ready_for_concurrent_call_waker: _,
            event_loop_running: _,
            #[cfg(feature = "task-group-hook")]
                task_group_hook: _,
        } = self;

        for (_, entry) in table.get_mut().iter_mut() {
            if let Some(set) = entry.downcast_mut::<WaitableSet>() {
                for mode in set.waiting.values_mut() {
                    match mode {
                        WaitMode::Fiber(fiber) => {
                            fiber.trace_gc_roots(modules, unwind, gc_roots_list);
                        }
                        WaitMode::Callback(_) => {}
                    }
                }
            } else if let Some(thread) = entry.downcast_mut::<GuestThread>() {
                match &mut thread.state {
                    GuestThreadState::Suspended(fiber) | GuestThreadState::Ready { fiber } => {
                        fiber.trace_gc_roots(modules, unwind, gc_roots_list);
                    }
                    GuestThreadState::NotStartedImplicit
                    | GuestThreadState::NotStartedExplicit(_)
                    | GuestThreadState::Running
                    | GuestThreadState::Completed => {}
                }
            } else if let Some(Some(WorkItem::ResumeFiber { fiber, .. })) =
                entry.downcast_mut::<Option<WorkItem>>()
            {
                fiber.trace_gc_roots(modules, unwind, gc_roots_list);
            }
        }

        if let Some(fiber) = worker {
            fiber.trace_gc_roots(modules, unwind, gc_roots_list);
        }

        let mut handle_item = |item: &mut WorkItem| match item {
            WorkItem::ResumeFiber { fiber, .. } => {
                fiber.trace_gc_roots(modules, unwind, gc_roots_list);
            }
            WorkItem::PushFuture(_future) => {
                // TODO(cm-gc): once futures can contain GC roots, we will need
                // to trace them.
            }
            WorkItem::ResumeThread { .. }
            | WorkItem::GuestCall { .. }
            | WorkItem::WorkerFunction(_) => {}
        };

        if let Some(item) = switch_item {
            handle_item(item);
        }
        if let Some(item) = next_switch_item {
            handle_item(item);
        }
        for item in high_priority {
            handle_item(item);
        }
        for item in low_priority {
            handle_item(item);
        }
        for item in saved_next_switch_items
            .iter_mut()
            .filter_map(|v| v.as_mut())
        {
            handle_item(item);
        }
    }

    fn schedule_suspended_thread(
        &mut self,
        thread: QualifiedThreadId,
        instance: RuntimeInstance,
        priority: Priority,
    ) -> Result<()> {
        // Validate the destination while the source still owns the fiber.
        if matches!(priority, Priority::Switch) && self.switch_item.is_some() {
            bail_bug!("switch item already set");
        }
        let slot = &mut self.get_mut(thread.thread)?.state;
        if !matches!(slot, GuestThreadState::Suspended(_)) {
            bail!(Trap::CannotResumeThread);
        }
        let GuestThreadState::Suspended(fiber) = mem::replace(slot, GuestThreadState::Running)
        else {
            unreachable!()
        };
        let item = WorkItem::ResumeFiber {
            instance,
            thread,
            fiber,
        };
        match priority {
            Priority::Switch => self.switch_item = Some(item),
            Priority::High => self.push_high_priority(item),
            Priority::Low => self.push_low_priority(item),
        }
        Ok(())
    }

    fn set_thread_running(&mut self, thread: TableId<GuestThread>) -> Result<()> {
        let slot = &mut self.get_mut(thread)?.state;
        slot.check_no_fiber()?;
        *slot = GuestThreadState::Running;
        Ok(())
    }

    fn save_next_switch_item(&mut self) -> Result<TableId<Option<WorkItem>>> {
        // push can consume and drop its argument when the table is full.
        // Reserve with an empty item before taking the fiber-owning source.
        let saved = self.push(None::<WorkItem>)?;
        let slot = self.table.get_mut().get_mut(&Resource::from(saved))?;
        *slot = self.next_switch_item.take();
        Ok(saved)
    }

    fn restore_next_switch_item(&mut self, saved: TableId<Option<WorkItem>>) -> Result<()> {
        if self.next_switch_item.is_some() {
            bail_bug!("next switch item already set when restoring");
        }
        self.next_switch_item = self.delete(saved)?;
        Ok(())
    }

    fn take_callback_waiter(
        &mut self,
        set: TableId<WaitableSet>,
        thread: QualifiedThreadId,
    ) -> Result<Instance> {
        let waiting = &mut self.get_mut(set)?.waiting;
        let instance = match waiting.get(&thread) {
            Some(WaitMode::Callback(instance)) => *instance,
            other => bail_bug!("expected `Some(WaitMode::Callback(_))`; got `{other:?}`"),
        };
        waiting.remove(&thread);
        Ok(instance)
    }

    fn push<V: Send + Sync + 'static>(
        &mut self,
        value: V,
    ) -> Result<TableId<V>, ResourceTableError> {
        self.table.get_mut().push(value).map(TableId::from)
    }

    fn get_mut<V: 'static>(&mut self, id: TableId<V>) -> Result<&mut V, ResourceTableError> {
        self.table.get_mut().get_mut(&Resource::from(id))
    }

    pub fn add_child<T: 'static, U: 'static>(
        &mut self,
        child: TableId<T>,
        parent: TableId<U>,
    ) -> Result<(), ResourceTableError> {
        self.table
            .get_mut()
            .add_child(Resource::from(child), Resource::from(parent))
    }

    pub fn remove_child<T: 'static, U: 'static>(
        &mut self,
        child: TableId<T>,
        parent: TableId<U>,
    ) -> Result<(), ResourceTableError> {
        self.table
            .get_mut()
            .remove_child(Resource::from(child), Resource::from(parent))
    }

    fn delete<V: 'static>(&mut self, id: TableId<V>) -> Result<V, ResourceTableError> {
        self.table.get_mut().delete(Resource::from(id))
    }

    fn push_future(&mut self, future: HostTaskFuture) {
        // Note that we can't directly push to `ConcurrentState::futures` here
        // since this may be called from a future that's being polled inside
        // `Self::poll_until`, which temporarily removes the `FuturesUnordered`
        // so it has exclusive access while polling it.  Therefore, we push a
        // work item to the "high priority" queue, which will actually push to
        // `ConcurrentState::futures` later.
        self.push_high_priority(WorkItem::PushFuture(AlwaysMut::new(future)));
    }

    fn set_switch_item(&mut self, item: WorkItem) -> Result<()> {
        if self.switch_item.is_some() {
            // Preserve ownership on this internal-error path. The caller
            // must stop scheduling after the error; teardown disposes the item.
            self.push_high_priority(item);
            bail_bug!("switch item already set");
        }

        self.switch_item = Some(item);
        log::trace!("set switch item: {:?}", self.switch_item.as_ref().unwrap());

        Ok(())
    }

    fn take_next_switch_item(&mut self) -> Result<()> {
        if let Some(item) = self.next_switch_item.take() {
            self.set_switch_item(item)?;
        }
        Ok(())
    }

    /// Wake the first thread waiting on `set`, if any, so that it may receive
    /// one of the set's pending events.
    fn wake_waiter(&mut self, set: TableId<WaitableSet>) -> Result<()> {
        // Finish every fallible lookup and check before taking the waiter out
        // of the set, so that an error leaves its fiber owned by the set.
        let Some((&thread, _)) = self.get_mut(set)?.waiting.first_key_value() else {
            return Ok(());
        };
        let wake_on_cancel = self.get_mut(thread.thread)?.wake_on_cancel;
        if wake_on_cancel != WakeOnCancel::None && wake_on_cancel != WakeOnCancel::Waiting(set) {
            bail_bug!("thread {thread:?} has unexpected wake_on_cancel value {wake_on_cancel:?}");
        }
        let instance = self.get_mut(thread.task)?.instance;
        self.get_mut(thread.thread)?.wake_on_cancel.take();
        let (_, mode) = self.get_mut(set)?.waiting.pop_first().unwrap();

        let item = match mode {
            WaitMode::Fiber(fiber) => WorkItem::ResumeFiber {
                instance,
                thread,
                fiber,
            },
            WaitMode::Callback(callback_instance) => WorkItem::GuestCall {
                instance,
                call: GuestCall {
                    thread,
                    kind: GuestCallKind::DeliverEvent {
                        instance: callback_instance,
                        set: Some(set),
                    },
                },
            },
        };
        self.push_high_priority(item);
        Ok(())
    }

    fn push_high_priority(&mut self, item: WorkItem) {
        self.high_priority.push_front(item);
        log::trace!(
            "push high priority: {:?}",
            self.high_priority.front().unwrap()
        );
    }

    fn push_low_priority(&mut self, item: WorkItem) {
        self.low_priority.push_front(item);
        log::trace!(
            "push low priority: {:?}",
            self.low_priority.front().unwrap()
        );
    }

    fn push_work_item(&mut self, item: WorkItem, priority: Priority) -> Result<()> {
        match priority {
            Priority::Switch => self.set_switch_item(item)?,
            Priority::High => self.push_high_priority(item),
            Priority::Low => self.push_low_priority(item),
        }

        Ok(())
    }

    fn promote_instance_local_thread_work_item(
        &mut self,
        current_instance: RuntimeInstance,
    ) -> Result<bool> {
        log::trace!("promote thread work items for {current_instance:?}");

        self.promote_work_item_matching(|item: &WorkItem| {
            let result = match item {
                WorkItem::ResumeThread { instance, .. }
                | WorkItem::ResumeFiber { instance, .. }
                | WorkItem::GuestCall { instance, .. } => *instance == current_instance,
                _ => false,
            };

            log::trace!("candidate {item:?}: {result}");
            result
        })
    }

    fn promote_thread_work_item(&mut self, thread: QualifiedThreadId) -> Result<bool> {
        self.promote_work_item_matching(|item: &WorkItem| match item {
            WorkItem::ResumeThread {
                thread: item_thread,
                ..
            }
            | WorkItem::GuestCall {
                call:
                    GuestCall {
                        thread: item_thread,
                        ..
                    },
                ..
            } => *item_thread == thread,
            _ => false,
        })
    }

    fn promote_work_item_matching<F>(&mut self, mut predicate: F) -> Result<bool>
    where
        F: FnMut(&WorkItem) -> bool,
    {
        // Scan from the back to preserve the ordering of items
        // are popped from the back of the `VecDeque`s by `poll_until` and
        // pushed to the front by `push_{high,low}_priority`.

        // Inspect items while they are still owned by the store. A predicate
        // (including tracing) may panic; it must not drop an extracted queue.
        if self.switch_item.is_none() {
            if let Some(index) = self.high_priority.iter().rposition(&mut predicate) {
                self.switch_item = self.high_priority.remove(index);
            }
        }
        if self.switch_item.is_none() {
            if let Some(index) = self.low_priority.iter().rposition(predicate) {
                self.switch_item = self.low_priority.remove(index);
            }
        }
        Ok(self.switch_item.is_some())
    }

    /// Used by `ResourceTables` to acquire the current `CallContext` for the
    /// specified task.
    pub fn call_context(&mut self, task: Scope) -> Result<&mut CallContext> {
        match task {
            Scope::HostId(task) => {
                let task: TableId<HostTask> = TableId::new(task);
                Ok(&mut self.get_mut(task)?.call_context)
            }
            Scope::Id(task) => {
                let task: TableId<GuestTask> = TableId::new(task);
                Ok(&mut self.get_mut(task)?.call_context)
            }
        }
    }

    pub(crate) fn deferred_host_call_context(&mut self) -> Option<&mut CallContext> {
        self.deferred_host_call_context.as_mut()
    }

    fn futures_mut(&mut self) -> Result<&mut FuturesUnordered<HostTaskFuture>> {
        match self.futures.get_mut().as_mut() {
            Some(f) => Ok(f),
            None => bail_bug!("futures field of concurrent state is currently taken"),
        }
    }

    pub(crate) fn table(&mut self) -> &mut ResourceTable {
        self.table.get_mut()
    }

    fn debug_assert_deferred_host_invariant(&self) {
        debug_assert_eq!(
            self.deferred_host_call_context.is_some(),
            matches!(self.unforced_current_thread, CurrentThread::DeferredHost(_)),
            "a deferred host thread and call context must exist together",
        );
    }

    fn materialize_host_task(&mut self) -> Result<CurrentThread> {
        self.debug_assert_deferred_host_invariant();
        let caller = match self.unforced_current_thread {
            CurrentThread::DeferredHost(caller) => caller,
            thread => return Ok(thread),
        };

        // Push first so allocation failure leaves the deferred state intact.
        let task = HostTask::new(self, HostTaskState::CalleeStarted, caller)?;
        let task = self.push(task)?;
        let call_context = self
            .deferred_host_call_context
            .take()
            .expect("deferred host call context should be present");
        self.get_mut(task)
            .expect("newly inserted host task should be present")
            .call_context = call_context;
        self.unforced_current_thread = CurrentThread::Host(task);
        self.debug_assert_deferred_host_invariant();
        log::trace!("new host task materialized {task:?}");
        Ok(CurrentThread::Host(task))
    }

    fn materialize_current_host_task_id(&mut self) -> Result<Option<TableId<HostTask>>> {
        match self.materialize_host_task()? {
            CurrentThread::Host(id) => Ok(Some(id)),
            CurrentThread::None => Ok(None),
            CurrentThread::Guest(_) => {
                bail_bug!("tried to materialize a host task id from a guest thread")
            }
            CurrentThread::DeferredHost(_) => {
                bail_bug!(
                    "current thread is a deferred host thread which should have been materialized"
                )
            }
        }
    }

    pub(crate) fn materialize_current_scope(&mut self) -> Result<Scope> {
        match self.materialize_host_task()? {
            CurrentThread::Host(id) => Ok(Scope::HostId(id.rep())),
            _ => bail_bug!("current scope is not a deferred host scope"),
        }
    }

    fn increment_task_ref_count(&mut self, task: TableId<GuestTask>) -> Result<()> {
        let count = &mut self.get_mut(task)?.ref_count;
        *count += 1;
        log::trace!("increment {task:?} to {count}");
        Ok(())
    }

    fn decrement_task_ref_count(&mut self, task: TableId<GuestTask>) -> Result<()> {
        let count = &mut self.get_mut(task)?.ref_count;
        if *count == 0 {
            bail_bug!("can't decrement task ref count that's already zero");
        }
        *count -= 1;
        log::trace!("decrement {task:?} to {count}");
        if *count == 0 {
            let task = self.delete(task)?;

            self.decrement_group_ref_count(task.group)?;

            // When a guest task is created it increments the
            // `ConcurrentState::interesting_tasks` counter, and that needs
            // to be paired with a decrement. There are a few situations in
            // which the decrement needs to happen which don't all funnel
            // through here, so in lieu of that at least try to catch issues
            // where we forgot to do a decrement.
            debug_assert!(task.decremented_interesting_task_count);
        }
        Ok(())
    }

    fn delete_host_task(&mut self, task: TableId<HostTask>) -> Result<()> {
        log::trace!("delete host task {task:?}");
        let task = self.delete(task)?;

        self.decrement_group_ref_count(task.group)
    }
}

/// Provide a type hint to compiler about the shape of a parameter lower
/// closure.
fn for_any_lower<
    F: FnOnce(&mut dyn VMStore, &mut [MaybeUninit<ValRaw>]) -> Result<()> + Send + Sync,
>(
    fun: F,
) -> F {
    fun
}

/// Provide a type hint to compiler about the shape of a result lift closure.
fn for_any_lift<
    F: FnOnce(&mut dyn VMStore, &[ValRaw]) -> Result<Box<dyn Any + Send + Sync>> + Send + Sync,
>(
    fun: F,
) -> F {
    fun
}

fn check_ambient_store(id: StoreId) {
    let message = "\
        `Future`s which depend on asynchronous component tasks, streams, or \
        futures to complete may only be polled from the event loop of the \
        store to which they belong.  Please use \
        `StoreContextMut::{run_concurrent,spawn}` to poll or await them.\
    ";
    tls::try_get(|store| {
        let matched = match store {
            tls::TryGet::Some(store) => store.id() == id,
            tls::TryGet::Taken | tls::TryGet::None => false,
        };

        if !matched {
            panic!("{message}")
        }
    });
}

fn unpack_callback_code(code: u32) -> (u32, u32) {
    (code & 0xF, code >> 4)
}

/// Helper struct for packaging parameters to be passed to
/// `ComponentInstance::waitable_check` for calls to `waitable-set.wait` or
/// `waitable-set.poll`.
struct WaitableCheckParams {
    set: TableId<WaitableSet>,
    options: OptionsIndex,
    payload: u32,
}

/// Indicates whether `ComponentInstance::waitable_check` is being called for
/// `waitable-set.wait` or `waitable-set.poll`.
enum WaitableCheck {
    Wait,
    Poll,
}

/// Represents a guest task called from the host, prepared using `prepare_call`.
pub(crate) struct PreparedCall<R> {
    /// The guest export to be called
    handle: Func,
    /// The guest thread created by `prepare_call`
    thread: QualifiedThreadId,
    /// The number of lowered core Wasm parameters to pass to the call.
    param_count: usize,
    /// The `oneshot::Receiver` to which the result of the call will be
    /// delivered when it is available.
    rx: oneshot::Receiver<LiftedResult>,
    /// The instance that this call is prepared for.
    runtime_instance: RuntimeInstance,
    _phantom: PhantomData<R>,
}

impl<R> PreparedCall<R> {
    /// Acquire a reference to the `TaskId` for this `PreparedCall`.
    ///
    /// Note that this increments the reference count for the task, and
    /// `TaskId::host_future_dropped` decrements it.  Hence,
    /// `TaskId::host_future_dropped` _must_ be called before the `TaskId` is
    /// dropped to avoid leaking the task.
    pub(crate) fn acquire_task_id(&self, store: &mut StoreOpaque) -> Result<TaskId> {
        store
            .concurrent_state_mut()?
            .increment_task_ref_count(self.thread.task)?;

        Ok(TaskId {
            task: self.thread.task,
            runtime_instance: self.runtime_instance,
        })
    }
}

/// Represents a task created by `prepare_call`.
pub(crate) struct TaskId {
    task: TableId<GuestTask>,
    runtime_instance: RuntimeInstance,
}

impl TaskId {
    /// The host future for an async task was dropped. If the parameters have
    /// not been lowered yet, it is no longer valid to do so, as the lowering
    /// closure would see a dangling pointer. In this case, we clear
    /// `GuestTask::lower_params` and cancel the task to ensure that doesn't
    /// happen. If the parameters _have_ already been lowered, then the closure
    /// is no longer reachable, in which case we only need to decrement the task
    /// refcount.
    pub(crate) fn host_future_dropped(&self, store: &mut StoreOpaque) -> Result<()> {
        let task = store.concurrent_state_mut()?.get_mut(self.task)?;
        if !task.already_lowered_parameters() {
            store.cancel_guest_subtask_without_lowered_parameters(
                self.runtime_instance,
                self.task,
            )?;
        }

        store
            .concurrent_state_mut()?
            .decrement_task_ref_count(self.task)?;

        Ok(())
    }
}

/// Prepare a call to the specified exported Wasm function, providing functions
/// for lowering the parameters and lifting the result.
///
/// To enqueue the returned `PreparedCall` in the `ComponentInstance`'s event
/// loop, use `stage_call`.
pub(crate) fn prepare_call<T, R>(
    mut store: StoreContextMut<T>,
    handle: Func,
    param_count: usize,
    lower_params: impl FnOnce(StoreContextMut<T>, &mut [MaybeUninit<ValRaw>]) -> Result<()>
    + Send
    + Sync
    + 'static,
    lift_result: impl FnOnce(&mut StoreOpaque, &[ValRaw]) -> Result<Box<dyn Any + Send + Sync>>
    + Send
    + Sync
    + 'static,
) -> Result<PreparedCall<R>> {
    if !store.0.may_enter() {
        bail!(Trap::CannotEnterComponent);
    }

    let (options, _flags, ty, raw_options) = handle.abi_info(store.0);

    let instance = handle.instance().id().get(store.0);
    let options = &instance.component().env_component().options[options];
    let ty = &instance.component().types()[ty];
    let async_typed = ty.async_;
    let async_lifted = raw_options.async_;
    let task_return_type = ty.results;
    let component_instance = raw_options.instance;
    let callback = options.callback.map(|i| instance.runtime_callback(i));
    let memory = options
        .memory()
        .map(|i| instance.runtime_memory(i))
        .map(SendSyncPtr::new);
    let string_encoding = options.string_encoding;
    let token = StoreToken::new(store.as_context_mut());
    let caller = store.0.materialize_host_task_id()?;
    let state = store.0.concurrent_state_mut()?;

    let (tx, rx) = oneshot::channel();

    let instance = handle.instance().runtime_instance(component_instance);
    let thread = GuestTask::new(
        state,
        Box::new(for_any_lower(move |store, params| {
            lower_params(token.as_context_mut(store), params)
        })),
        LiftResult {
            lift: Box::new(for_any_lift(move |store, result| {
                lift_result(store, result)
            })),
            ty: task_return_type,
            memory,
            string_encoding,
        },
        Caller::Host {
            tx: Some(tx),
            caller,
        },
        callback.map(|callback| {
            let callback = SendSyncPtr::new(callback);
            let instance = handle.instance();
            Box::new(move |store: &mut dyn VMStore, event, handle| {
                let store = token.as_context_mut(store);
                // SAFETY: Per the contract of `prepare_call`, the callback
                // will remain valid at least as long is this task exists.
                unsafe { instance.call_callback(store, callback, event, handle) }
            }) as CallbackFn
        }),
        instance,
        async_typed,
        async_lifted,
    )?;

    Ok(PreparedCall {
        handle,
        thread,
        param_count,
        runtime_instance: instance,
        rx,
        _phantom: PhantomData,
    })
}

pub(crate) struct StagedCall<R> {
    store: StoreId,
    rx: oneshot::Receiver<LiftedResult>,
    _marker: PhantomData<fn() -> R>,
    /// The task group to which the underlying task belongs.
    ///
    /// Note that this must be treated as opaque and not used to look up the
    /// corresponding `TaskGroup` since we decrement the task refcount in
    /// `StagedCall::new` which may cause the group to be disposed.  This field
    /// is only present to enable `[Typed]FuncCallConcurrent::group`, allowing
    /// the embedder to correlate the task group with `TaskGroupHook` events.
    group: TaskGroupId,
}

impl<R> StagedCall<R> {
    /// Queue a call previously prepared using `prepare_call` to be run as part of
    /// the associated `ComponentInstance`'s event loop.
    ///
    /// Note that this has the side-effect of disposing the specified
    /// `PrepareCall` and decrementing the task reference count since
    /// `StagedCall` has no further need to access the task directly.
    ///
    /// The returned future will resolve to the result once it is available, but
    /// must only be polled via the instance's event loop. See
    /// `StoreContextMut::run_concurrent` for details.
    pub(crate) fn new<T: 'static>(
        mut store: StoreContextMut<T>,
        prepared: PreparedCall<R>,
    ) -> Result<StagedCall<R>> {
        let PreparedCall {
            handle,
            thread,
            param_count,
            rx,
            ..
        } = prepared;

        stage_call0(store.as_context_mut(), handle, thread, param_count)?;

        let state = store.0.concurrent_state_mut()?;
        let group = state.get_mut(thread.task)?.group;
        state.decrement_task_ref_count(thread.task)?;

        Ok(StagedCall {
            store: store.0.id(),
            rx,
            _marker: PhantomData,
            group,
        })
    }
}

impl<R> Future for StagedCall<R>
where
    R: 'static,
{
    type Output = Result<R>;

    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
        check_ambient_store(self.store);
        Pin::new(&mut self.rx).poll(cx).map(|result| match result {
            Ok(r) => match r.downcast() {
                Ok(r) => Ok(*r),
                Err(_) => bail_bug!("wrong type of value produced"),
            },
            Err(oneshot::Canceled) => bail_bug!("channel erroneously dropped"),
        })
    }
}

/// Queue a call previously prepared using `prepare_call` to be run as part of
/// the associated `ComponentInstance`'s event loop.
fn stage_call0<T: 'static>(
    store: StoreContextMut<T>,
    handle: Func,
    guest_thread: QualifiedThreadId,
    param_count: usize,
) -> Result<()> {
    let (_options, _, _ty, raw_options) = handle.abi_info(store.0);
    let is_concurrent = raw_options.async_;
    let callback = raw_options.callback;
    let instance = handle.instance();
    let callee = handle.lifted_core_func(store.0);
    let post_return = raw_options
        .post_return
        .map(|i| instance.id().get(store.0).runtime_post_return(i));
    let callback = callback.map(|i| {
        let instance = instance.id().get(store.0);
        SendSyncPtr::new(instance.runtime_callback(i))
    });

    log::trace!("queueing call {guest_thread:?}");

    // SAFETY: `callee`, `callback`, and `post_return` are valid pointers
    // (with signatures appropriate for this call) and will remain valid as
    // long as this instance is valid.
    unsafe {
        instance.stage_call(
            store,
            guest_thread,
            SendSyncPtr::new(callee),
            param_count,
            1,
            is_concurrent,
            callback,
            post_return.map(SendSyncPtr::new),
            true,
        )
    }
}

#[cfg(all(test, feature = "cranelift", feature = "wat"))]
mod tests {
    use super::*;
    use crate::component::{Component, Linker};
    use crate::store::AsStoreOpaque;
    use crate::{Config, Engine};

    fn host_subtask(
        state: HostTaskState,
        event: Option<Event>,
    ) -> Result<(Store<()>, Instance, TableId<HostTask>, u32)> {
        let mut config = Config::new();
        config.wasm_component_model_async(true);
        let engine = Engine::new(&config)?;
        let component = Component::new(&engine, "(component)")?;
        let mut store = Store::new(&engine, ());
        let instance = Linker::new(&engine).instantiate(&mut store, &component)?;
        let store_opaque = store.as_store_opaque();
        let concurrent_state = store_opaque.concurrent_state_mut()?;
        // These tests create an isolated host task without a guest caller, so
        // give it its own task group instead of inheriting the caller's group.
        let group = concurrent_state.make_task_group()?;
        let task = concurrent_state.push(HostTask {
            common: WaitableCommon::default(),
            call_context: CallContext::default(),
            state,
            group,
        })?;
        let handle = store_opaque
            .instance_state(instance.runtime_instance(RuntimeComponentInstanceIndex::from_u32(0)))
            .handle_table()
            .subtask_insert_host(task.rep())?;
        let common = &mut store_opaque.concurrent_state_mut()?.get_mut(task)?.common;
        common.handle = Some(handle);
        common.event = event;
        Ok((store, instance, task, handle))
    }

    #[test]
    fn host_subtask_drop_during_cancellation() -> Result<()> {
        for abort_completed in [false, true] {
            let (handle, future) = JoinHandle::run(future::pending::<()>());
            let mut future = pin!(future);
            let (mut store, instance, task, handle) =
                host_subtask(HostTaskState::CalleeRunning(handle), None)?;
            let store = store.as_store_opaque();
            let caller = RuntimeComponentInstanceIndex::from_u32(0);
            assert_eq!(
                instance.subtask_cancel(store, caller, true, handle)?,
                BLOCKED
            );
            if abort_completed {
                // Even after the abort resolves, the completion worker still
                // needs the task entry to lower the result and publish an event.
                assert!(matches!(
                    future
                        .as_mut()
                        .poll(&mut Context::from_waker(Waker::noop())),
                    Poll::Ready(None),
                ));
            }
            for async_ in [false, true] {
                let err = instance
                    .subtask_cancel(store, caller, async_, handle)
                    .unwrap_err();
                assert_eq!(err.downcast::<Trap>()?, Trap::SubtaskCancelAfterTerminal);
            }
            let err = instance.subtask_drop(store, caller, handle).unwrap_err();
            assert_eq!(err.downcast::<Trap>()?, Trap::SubtaskDropNotResolved);
            assert!(store.concurrent_state_mut()?.get_mut(task).is_ok());
        }
        Ok(())
    }

    #[test]
    fn host_subtask_cancel_after_completion() -> Result<()> {
        for async_ in [false, true] {
            let (mut store, instance, task, handle) = host_subtask(
                HostTaskState::CalleeDone { cancelled: false },
                Some(Event::Subtask {
                    status: Status::Returned,
                }),
            )?;
            let store = store.as_store_opaque();
            let caller = RuntimeComponentInstanceIndex::from_u32(0);
            assert_eq!(
                instance.subtask_cancel(store, caller, async_, handle)?,
                Status::Returned as u32,
            );
            let err = instance
                .subtask_cancel(store, caller, async_, handle)
                .unwrap_err();
            assert_eq!(err.downcast::<Trap>()?, Trap::SubtaskCancelAfterTerminal);
            instance.subtask_drop(store, caller, handle)?;
            assert!(store.concurrent_state_mut()?.get_mut(task).is_err());
        }
        Ok(())
    }

    #[test]
    fn host_subtask_drop_requires_terminal_event_delivery() -> Result<()> {
        for (cancelled, status) in [
            (false, Status::Returned),
            (true, Status::Returned),
            (true, Status::ReturnCancelled),
        ] {
            for delivered in [false, true] {
                let event = if delivered {
                    None
                } else {
                    Some(Event::Subtask { status })
                };
                let (mut store, instance, task, handle) =
                    host_subtask(HostTaskState::CalleeDone { cancelled }, event)?;
                let store = store.as_store_opaque();
                let result = instance.subtask_drop(
                    store,
                    RuntimeComponentInstanceIndex::from_u32(0),
                    handle,
                );
                if delivered {
                    result?;
                    assert!(store.concurrent_state_mut()?.get_mut(task).is_err());
                } else {
                    let err = result.unwrap_err();
                    assert_eq!(err.downcast::<Trap>()?, Trap::SubtaskDropNotResolved);
                    assert!(store.concurrent_state_mut()?.get_mut(task).is_ok());
                }
            }
        }
        Ok(())
    }
}