regex_automata/util/

pool.rs

// This module provides a relatively simple thread-safe pool of reusable
// objects. For the most part, it's implemented by a stack represented by a
// Mutex<Vec<T>>. It has one small trick: because unlocking a mutex is somewhat
// costly, in the case where a pool is accessed by the first thread that tried
// to get a value, we bypass the mutex. Here are some benchmarks showing the
// difference.
//
// 2022-10-15: These benchmarks are from the old regex crate and they aren't
// easy to reproduce because some rely on older implementations of Pool that
// are no longer around. I've left the results here for posterity, but any
// enterprising individual should feel encouraged to re-litigate the way Pool
// works. I am not at all certain it is the best approach.
//
// 1) misc::anchored_literal_long_non_match    21 (18571 MB/s)
// 2) misc::anchored_literal_long_non_match   107 (3644 MB/s)
// 3) misc::anchored_literal_long_non_match    45 (8666 MB/s)
// 4) misc::anchored_literal_long_non_match    19 (20526 MB/s)
//
// (1) represents our baseline: the master branch at the time of writing when
// using the 'thread_local' crate to implement the pool below.
//
// (2) represents a naive pool implemented completely via Mutex<Vec<T>>. There
// is no special trick for bypassing the mutex.
//
// (3) is the same as (2), except it uses Mutex<Vec<Box<T>>>. It is twice as
// fast because a Box<T> is much smaller than the T we use with a Pool in this
// crate. So pushing and popping a Box<T> from a Vec is quite a bit faster
// than for T.
//
// (4) is the same as (3), but with the trick for bypassing the mutex in the
// case of the first-to-get thread.
//
// Why move off of thread_local? Even though (4) is a hair faster than (1)
// above, this was not the main goal. The main goal was to move off of
// thread_local and find a way to *simply* re-capture some of its speed for
// regex's specific case. So again, why move off of it? The *primary* reason is
// because of memory leaks. See https://github.com/rust-lang/regex/issues/362
// for example. (Why do I want it to be simple? Well, I suppose what I mean is,
// "use as much safe code as possible to minimize risk and be as sure as I can
// be that it is correct.")
//
// My guess is that the thread_local design is probably not appropriate for
// regex since its memory usage scales to the number of active threads that
// have used a regex, where as the pool below scales to the number of threads
// that simultaneously use a regex. While neither case permits contraction,
// since we own the pool data structure below, we can add contraction if a
// clear use case pops up in the wild. More pressingly though, it seems that
// there are at least some use case patterns where one might have many threads
// sitting around that might have used a regex at one point. While thread_local
// does try to reuse space previously used by a thread that has since stopped,
// its maximal memory usage still scales with the total number of active
// threads. In contrast, the pool below scales with the total number of threads
// *simultaneously* using the pool. The hope is that this uses less memory
// overall. And if it doesn't, we can hopefully tune it somehow.
//
// It seems that these sort of conditions happen frequently
// in FFI inside of other more "managed" languages. This was
// mentioned in the issue linked above, and also mentioned here:
// https://github.com/BurntSushi/rure-go/issues/3. And in particular, users
// confirm that disabling the use of thread_local resolves the leak.
//
// There were other weaker reasons for moving off of thread_local as well.
// Namely, at the time, I was looking to reduce dependencies. And for something
// like regex, maintenance can be simpler when we own the full dependency tree.
//
// Note that I am not entirely happy with this pool. It has some subtle
// implementation details and is overall still observable (even with the
// thread owner optimization) in benchmarks. If someone wants to take a crack
// at building something better, please file an issue. Even if it means a
// different API. The API exposed by this pool is not the minimal thing that
// something like a 'Regex' actually needs. It could adapt to, for example,
// an API more like what is found in the 'thread_local' crate. However, we do
// really need to support the no-std alloc-only context, or else the regex
// crate wouldn't be able to support no-std alloc-only. However, I'm generally
// okay with making the alloc-only context slower (as it is here), although I
// do find it unfortunate.

/*!
A thread safe memory pool.

The principal type in this module is a [`Pool`]. It main use case is for
holding a thread safe collection of mutable scratch spaces (usually called
`Cache` in this crate) that regex engines need to execute a search. This then
permits sharing the same read-only regex object across multiple threads while
having a quick way of reusing scratch space in a thread safe way. This avoids
needing to re-create the scratch space for every search, which could wind up
being quite expensive.
*/

/// A thread safe pool that works in an `alloc`-only context.
///
/// Getting a value out comes with a guard. When that guard is dropped, the
/// value is automatically put back in the pool. The guard provides both a
/// `Deref` and a `DerefMut` implementation for easy access to an underlying
/// `T`.
///
/// A `Pool` impls `Sync` when `T` is `Send` (even if `T` is not `Sync`). This
/// is possible because a pool is guaranteed to provide a value to exactly one
/// thread at any time.
///
/// Currently, a pool never contracts in size. Its size is proportional to the
/// maximum number of simultaneous uses. This may change in the future.
///
/// A `Pool` is a particularly useful data structure for this crate because
/// many of the regex engines require a mutable "cache" in order to execute
/// a search. Since regexes themselves tend to be global, the problem is then:
/// how do you get a mutable cache to execute a search? You could:
///
/// 1. Use a `thread_local!`, which requires the standard library and requires
/// that the regex pattern be statically known.
/// 2. Use a `Pool`.
/// 3. Make the cache an explicit dependency in your code and pass it around.
/// 4. Put the cache state in a `Mutex`, but this means only one search can
/// execute at a time.
/// 5. Create a new cache for every search.
///
/// A `thread_local!` is perhaps the best choice if it works for your use case.
/// Putting the cache in a mutex or creating a new cache for every search are
/// perhaps the worst choices. Of the remaining two choices, whether you use
/// this `Pool` or thread through a cache explicitly in your code is a matter
/// of taste and depends on your code architecture.
///
/// # Warning: may use a spin lock
///
/// When this crate is compiled _without_ the `std` feature, then this type
/// may used a spin lock internally. This can have subtle effects that may
/// be undesirable. See [Spinlocks Considered Harmful][spinharm] for a more
/// thorough treatment of this topic.
///
/// [spinharm]: https://matklad.github.io/2020/01/02/spinlocks-considered-harmful.html
///
/// # Example
///
/// This example shows how to share a single hybrid regex among multiple
/// threads, while also safely getting exclusive access to a hybrid's
/// [`Cache`](crate::hybrid::regex::Cache) without preventing other searches
/// from running while your thread uses the `Cache`.
///
/// ```
/// use regex_automata::{
///     hybrid::regex::{Cache, Regex},
///     util::{lazy::Lazy, pool::Pool},
///     Match,
/// };
///
/// static RE: Lazy<Regex> =
///     Lazy::new(|| Regex::new("foo[0-9]+bar").unwrap());
/// static CACHE: Lazy<Pool<Cache>> =
///     Lazy::new(|| Pool::new(|| RE.create_cache()));
///
/// let expected = Some(Match::must(0, 3..14));
/// assert_eq!(expected, RE.find(&mut CACHE.get(), b"zzzfoo12345barzzz"));
/// ```
pub struct Pool<T, F = fn() -> T>(alloc::boxed::Box<inner::Pool<T, F>>);

impl<T, F> Pool<T, F> {
    /// Create a new pool. The given closure is used to create values in
    /// the pool when necessary.
    pub fn new(create: F) -> Pool<T, F> {
        Pool(alloc::boxed::Box::new(inner::Pool::new(create)))
    }
}

impl<T: Send, F: Fn() -> T> Pool<T, F> {
    /// Get a value from the pool. The caller is guaranteed to have
    /// exclusive access to the given value. Namely, it is guaranteed that
    /// this will never return a value that was returned by another call to
    /// `get` but was not put back into the pool.
    ///
    /// When the guard goes out of scope and its destructor is called, then
    /// it will automatically be put back into the pool. Alternatively,
    /// [`PoolGuard::put`] may be used to explicitly put it back in the pool
    /// without relying on its destructor.
    ///
    /// Note that there is no guarantee provided about which value in the
    /// pool is returned. That is, calling get, dropping the guard (causing
    /// the value to go back into the pool) and then calling get again is
    /// *not* guaranteed to return the same value received in the first `get`
    /// call.
    #[inline]
    pub fn get(&self) -> PoolGuard<'_, T, F> {
        PoolGuard(self.0.get())
    }
}

impl<T: core::fmt::Debug, F> core::fmt::Debug for Pool<T, F> {
    fn fmt(&self, f: &mut core::fmt::Formatter) -> core::fmt::Result {
        f.debug_tuple("Pool").field(&self.0).finish()
    }
}

/// A guard that is returned when a caller requests a value from the pool.
///
/// The purpose of the guard is to use RAII to automatically put the value
/// back in the pool once it's dropped.
pub struct PoolGuard<'a, T: Send, F: Fn() -> T>(inner::PoolGuard<'a, T, F>);

impl<'a, T: Send, F: Fn() -> T> PoolGuard<'a, T, F> {
    /// Consumes this guard and puts it back into the pool.
    ///
    /// This circumvents the guard's `Drop` implementation. This can be useful
    /// in circumstances where the automatic `Drop` results in poorer codegen,
    /// such as calling non-inlined functions.
    #[inline]
    pub fn put(this: PoolGuard<'_, T, F>) {
        inner::PoolGuard::put(this.0);
    }
}

impl<'a, T: Send, F: Fn() -> T> core::ops::Deref for PoolGuard<'a, T, F> {
    type Target = T;

    #[inline]
    fn deref(&self) -> &T {
        self.0.value()
    }
}

impl<'a, T: Send, F: Fn() -> T> core::ops::DerefMut for PoolGuard<'a, T, F> {
    #[inline]
    fn deref_mut(&mut self) -> &mut T {
        self.0.value_mut()
    }
}

impl<'a, T: Send + core::fmt::Debug, F: Fn() -> T> core::fmt::Debug
    for PoolGuard<'a, T, F>
{
    fn fmt(&self, f: &mut core::fmt::Formatter) -> core::fmt::Result {
        f.debug_tuple("PoolGuard").field(&self.0).finish()
    }
}

#[cfg(feature = "std")]
mod inner {
    use core::{
        cell::UnsafeCell,
        panic::{RefUnwindSafe, UnwindSafe},
        sync::atomic::{AtomicUsize, Ordering},
    };

    use alloc::{boxed::Box, vec, vec::Vec};

    use std::{sync::Mutex, thread_local};

    /// An atomic counter used to allocate thread IDs.
    ///
    /// We specifically start our counter at 3 so that we can use the values
    /// less than it as sentinels.
    static COUNTER: AtomicUsize = AtomicUsize::new(3);

    /// A thread ID indicating that there is no owner. This is the initial
    /// state of a pool. Once a pool has an owner, there is no way to change
    /// it.
    static THREAD_ID_UNOWNED: usize = 0;

    /// A thread ID indicating that the special owner value is in use and not
    /// available. This state is useful for avoiding a case where the owner
    /// of a pool calls `get` before putting the result of a previous `get`
    /// call back into the pool.
    static THREAD_ID_INUSE: usize = 1;

    /// This sentinel is used to indicate that a guard has already been dropped
    /// and should not be re-dropped. We use this because our drop code can be
    /// called outside of Drop and thus there could be a bug in the internal
    /// implementation that results in trying to put the same guard back into
    /// the same pool multiple times, and *that* could result in UB if we
    /// didn't mark the guard as already having been put back in the pool.
    ///
    /// So this isn't strictly necessary, but this let's us define some
    /// routines as safe (like PoolGuard::put_imp) that we couldn't otherwise
    /// do.
    static THREAD_ID_DROPPED: usize = 2;

    /// The number of stacks we use inside of the pool. These are only used for
    /// non-owners. That is, these represent the "slow" path.
    ///
    /// In the original implementation of this pool, we only used a single
    /// stack. While this might be okay for a couple threads, the prevalence of
    /// 32, 64 and even 128 core CPUs has made it untenable. The contention
    /// such an environment introduces when threads are doing a lot of searches
    /// on short haystacks (a not uncommon use case) is palpable and leads to
    /// huge slowdowns.
    ///
    /// This constant reflects a change from using one stack to the number of
    /// stacks that this constant is set to. The stack for a particular thread
    /// is simply chosen by `thread_id % MAX_POOL_STACKS`. The idea behind
    /// this setup is that there should be a good chance that accesses to the
    /// pool will be distributed over several stacks instead of all of them
    /// converging to one.
    ///
    /// This is not a particularly smart or dynamic strategy. Fixing this to a
    /// specific number has at least two downsides. First is that it will help,
    /// say, an 8 core CPU more than it will a 128 core CPU. (But, crucially,
    /// it will still help the 128 core case.) Second is that this may wind
    /// up being a little wasteful with respect to memory usage. Namely, if a
    /// regex is used on one thread and then moved to another thread, then it
    /// could result in creating a new copy of the data in the pool even though
    /// only one is actually needed.
    ///
    /// And that memory usage bit is why this is set to 8 and not, say, 64.
    /// Keeping it at 8 limits, to an extent, how much unnecessary memory can
    /// be allocated.
    ///
    /// In an ideal world, we'd be able to have something like this:
    ///
    /// * Grow the number of stacks as the number of concurrent callers
    /// increases. I spent a little time trying this, but even just adding an
    /// atomic addition/subtraction for each pop/push for tracking concurrent
    /// callers led to a big perf hit. Since even more work would seemingly be
    /// required than just an addition/subtraction, I abandoned this approach.
    /// * The maximum amount of memory used should scale with respect to the
    /// number of concurrent callers and *not* the total number of existing
    /// threads. This is primarily why the `thread_local` crate isn't used, as
    /// as some environments spin up a lot of threads. This led to multiple
    /// reports of extremely high memory usage (often described as memory
    /// leaks).
    /// * Even more ideally, the pool should contract in size. That is, it
    /// should grow with bursts and then shrink. But this is a pretty thorny
    /// issue to tackle and it might be better to just not.
    /// * It would be nice to explore the use of, say, a lock-free stack
    /// instead of using a mutex to guard a `Vec` that is ultimately just
    /// treated as a stack. The main thing preventing me from exploring this
    /// is the ABA problem. The `crossbeam` crate has tools for dealing with
    /// this sort of problem (via its epoch based memory reclamation strategy),
    /// but I can't justify bringing in all of `crossbeam` as a dependency of
    /// `regex` for this.
    ///
    /// See this issue for more context and discussion:
    /// https://github.com/rust-lang/regex/issues/934
    const MAX_POOL_STACKS: usize = 8;

    thread_local!(
        /// A thread local used to assign an ID to a thread.
        static THREAD_ID: usize = {
            let next = COUNTER.fetch_add(1, Ordering::Relaxed);
            // SAFETY: We cannot permit the reuse of thread IDs since reusing a
            // thread ID might result in more than one thread "owning" a pool,
            // and thus, permit accessing a mutable value from multiple threads
            // simultaneously without synchronization. The intent of this panic
            // is to be a sanity check. It is not expected that the thread ID
            // space will actually be exhausted in practice. Even on a 32-bit
            // system, it would require spawning 2^32 threads (although they
            // wouldn't all need to run simultaneously, so it is in theory
            // possible).
            //
            // This checks that the counter never wraps around, since atomic
            // addition wraps around on overflow.
            if next == 0 {
                panic!("regex: thread ID allocation space exhausted");
            }
            next
        };
    );

    /// This puts each stack in the pool below into its own cache line. This is
    /// an absolutely critical optimization that tends to have the most impact
    /// in high contention workloads. Without forcing each mutex protected
    /// into its own cache line, high contention exacerbates the performance
    /// problem by causing "false sharing." By putting each mutex in its own
    /// cache-line, we avoid the false sharing problem and the affects of
    /// contention are greatly reduced.
    #[derive(Debug)]
    #[repr(C, align(64))]
    struct CacheLine<T>(T);

    /// A thread safe pool utilizing std-only features.
    ///
    /// The main difference between this and the simplistic alloc-only pool is
    /// the use of std::sync::Mutex and an "owner thread" optimization that
    /// makes accesses by the owner of a pool faster than all other threads.
    /// This makes the common case of running a regex within a single thread
    /// faster by avoiding mutex unlocking.
    pub(super) struct Pool<T, F> {
        /// A function to create more T values when stack is empty and a caller
        /// has requested a T.
        create: F,
        /// Multiple stacks of T values to hand out. These are used when a Pool
        /// is accessed by a thread that didn't create it.
        ///
        /// Conceptually this is `Mutex<Vec<Box<T>>>`, but sharded out to make
        /// it scale better under high contention work-loads. We index into
        /// this sequence via `thread_id % stacks.len()`.
        stacks: Vec<CacheLine<Mutex<Vec<Box<T>>>>>,
        /// The ID of the thread that owns this pool. The owner is the thread
        /// that makes the first call to 'get'. When the owner calls 'get', it
        /// gets 'owner_val' directly instead of returning a T from 'stack'.
        /// See comments elsewhere for details, but this is intended to be an
        /// optimization for the common case that makes getting a T faster.
        ///
        /// It is initialized to a value of zero (an impossible thread ID) as a
        /// sentinel to indicate that it is unowned.
        owner: AtomicUsize,
        /// A value to return when the caller is in the same thread that
        /// first called `Pool::get`.
        ///
        /// This is set to None when a Pool is first created, and set to Some
        /// once the first thread calls Pool::get.
        owner_val: UnsafeCell<Option<T>>,
    }

    // SAFETY: Since we want to use a Pool from multiple threads simultaneously
    // behind an Arc, we need for it to be Sync. In cases where T is sync,
    // Pool<T> would be Sync. However, since we use a Pool to store mutable
    // scratch space, we wind up using a T that has interior mutability and is
    // thus itself not Sync. So what we *really* want is for our Pool<T> to by
    // Sync even when T is not Sync (but is at least Send).
    //
    // The only non-sync aspect of a Pool is its 'owner_val' field, which is
    // used to implement faster access to a pool value in the common case of
    // a pool being accessed in the same thread in which it was created. The
    // 'stack' field is also shared, but a Mutex<T> where T: Send is already
    // Sync. So we only need to worry about 'owner_val'.
    //
    // The key is to guarantee that 'owner_val' can only ever be accessed from
    // one thread. In our implementation below, we guarantee this by only
    // returning the 'owner_val' when the ID of the current thread matches the
    // ID of the thread that first called 'Pool::get'. Since this can only ever
    // be one thread, it follows that only one thread can access 'owner_val' at
    // any point in time. Thus, it is safe to declare that Pool<T> is Sync when
    // T is Send.
    //
    // If there is a way to achieve our performance goals using safe code, then
    // I would very much welcome a patch. As it stands, the implementation
    // below tries to balance safety with performance. The case where a Regex
    // is used from multiple threads simultaneously will suffer a bit since
    // getting a value out of the pool will require unlocking a mutex.
    //
    // We require `F: Send + Sync` because we call `F` at any point on demand,
    // potentially from multiple threads simultaneously.
    unsafe impl<T: Send, F: Send + Sync> Sync for Pool<T, F> {}

    // If T is UnwindSafe, then since we provide exclusive access to any
    // particular value in the pool, the pool should therefore also be
    // considered UnwindSafe.
    //
    // We require `F: UnwindSafe + RefUnwindSafe` because we call `F` at any
    // point on demand, so it needs to be unwind safe on both dimensions for
    // the entire Pool to be unwind safe.
    impl<T: UnwindSafe, F: UnwindSafe + RefUnwindSafe> UnwindSafe for Pool<T, F> {}

    // If T is UnwindSafe, then since we provide exclusive access to any
    // particular value in the pool, the pool should therefore also be
    // considered RefUnwindSafe.
    //
    // We require `F: UnwindSafe + RefUnwindSafe` because we call `F` at any
    // point on demand, so it needs to be unwind safe on both dimensions for
    // the entire Pool to be unwind safe.
    impl<T: UnwindSafe, F: UnwindSafe + RefUnwindSafe> RefUnwindSafe
        for Pool<T, F>
    {
    }

    impl<T, F> Pool<T, F> {
        /// Create a new pool. The given closure is used to create values in
        /// the pool when necessary.
        pub(super) fn new(create: F) -> Pool<T, F> {
            // FIXME: Now that we require 1.65+, Mutex::new is available as
            // const... So we can almost mark this function as const. But of
            // course, we're creating a Vec of stacks below (we didn't when I
            // originally wrote this code). It seems like the best way to work
            // around this would be to use a `[Stack; MAX_POOL_STACKS]` instead
            // of a `Vec<Stack>`. I refrained from making this change at time
            // of writing (2023/10/08) because I was making a lot of other
            // changes at the same time and wanted to do this more carefully.
            // Namely, because of the cache line optimization, that `[Stack;
            // MAX_POOL_STACKS]` would be quite big. It's unclear how bad (if
            // at all) that would be.
            //
            // Another choice would be to lazily allocate the stacks, but...
            // I'm not so sure about that. Seems like a fair bit of complexity?
            //
            // Maybe there's a simple solution I'm missing.
            //
            // ... OK, I tried to fix this. First, I did it by putting `stacks`
            // in an `UnsafeCell` and using a `Once` to lazily initialize it.
            // I benchmarked it and everything looked okay. I then made this
            // function `const` and thought I was just about done. But the
            // public pool type wraps its inner pool in a `Box` to keep its
            // size down. Blech.
            //
            // So then I thought that I could push the box down into this
            // type (and leave the non-std version unboxed) and use the same
            // `UnsafeCell` technique to lazily initialize it. This has the
            // downside of the `Once` now needing to get hit in the owner fast
            // path, but maybe that's OK? However, I then realized that we can
            // only lazily initialize `stacks`, `owner` and `owner_val`. The
            // `create` function needs to be put somewhere outside of the box.
            // So now the pool is a `Box`, `Once` and a function. Now we're
            // starting to defeat the point of boxing in the first place. So I
            // backed out that change too.
            //
            // Back to square one. I maybe we just don't make a pool's
            // constructor const and live with it. It's probably not a huge
            // deal.
            let mut stacks = Vec::with_capacity(MAX_POOL_STACKS);
            for _ in 0..stacks.capacity() {
                stacks.push(CacheLine(Mutex::new(vec![])));
            }
            let owner = AtomicUsize::new(THREAD_ID_UNOWNED);
            let owner_val = UnsafeCell::new(None); // init'd on first access
            Pool { create, stacks, owner, owner_val }
        }
    }

    impl<T: Send, F: Fn() -> T> Pool<T, F> {
        /// Get a value from the pool. This may block if another thread is also
        /// attempting to retrieve a value from the pool.
        #[inline]
        pub(super) fn get(&self) -> PoolGuard<'_, T, F> {
            // Our fast path checks if the caller is the thread that "owns"
            // this pool. Or stated differently, whether it is the first thread
            // that tried to extract a value from the pool. If it is, then we
            // can return a T to the caller without going through a mutex.
            //
            // SAFETY: We must guarantee that only one thread gets access
            // to this value. Since a thread is uniquely identified by the
            // THREAD_ID thread local, it follows that if the caller's thread
            // ID is equal to the owner, then only one thread may receive this
            // value. This is also why we can get away with what looks like a
            // racy load and a store. We know that if 'owner == caller', then
            // only one thread can be here, so we don't need to worry about any
            // other thread setting the owner to something else.
            let caller = THREAD_ID.with(|id| *id);
            let owner = self.owner.load(Ordering::Acquire);
            if caller == owner {
                // N.B. We could also do a CAS here instead of a load/store,
                // but ad hoc benchmarking suggests it is slower. And a lot
                // slower in the case where `get_slow` is common.
                self.owner.store(THREAD_ID_INUSE, Ordering::Release);
                return self.guard_owned(caller);
            }
            self.get_slow(caller, owner)
        }

        /// This is the "slow" version that goes through a mutex to pop an
        /// allocated value off a stack to return to the caller. (Or, if the
        /// stack is empty, a new value is created.)
        ///
        /// If the pool has no owner, then this will set the owner.
        #[cold]
        fn get_slow(
            &self,
            caller: usize,
            owner: usize,
        ) -> PoolGuard<'_, T, F> {
            if owner == THREAD_ID_UNOWNED {
                // This sentinel means this pool is not yet owned. We try to
                // atomically set the owner. If we do, then this thread becomes
                // the owner and we can return a guard that represents the
                // special T for the owner.
                //
                // Note that we set the owner to a different sentinel that
                // indicates that the owned value is in use. The owner ID will
                // get updated to the actual ID of this thread once the guard
                // returned by this function is put back into the pool.
                let res = self.owner.compare_exchange(
                    THREAD_ID_UNOWNED,
                    THREAD_ID_INUSE,
                    Ordering::AcqRel,
                    Ordering::Acquire,
                );
                if res.is_ok() {
                    // SAFETY: A successful CAS above implies this thread is
                    // the owner and that this is the only such thread that
                    // can reach here. Thus, there is no data race.
                    unsafe {
                        *self.owner_val.get() = Some((self.create)());
                    }
                    return self.guard_owned(caller);
                }
            }
            let stack_id = caller % self.stacks.len();
            // We try to acquire exclusive access to this thread's stack, and
            // if so, grab a value from it if we can. We put this in a loop so
            // that it's easy to tweak and experiment with a different number
            // of tries. In the end, I couldn't see anything obviously better
            // than one attempt in ad hoc testing.
            for _ in 0..1 {
                let mut stack = match self.stacks[stack_id].0.try_lock() {
                    Err(_) => continue,
                    Ok(stack) => stack,
                };
                if let Some(value) = stack.pop() {
                    return self.guard_stack(value);
                }
                // Unlock the mutex guarding the stack before creating a fresh
                // value since we no longer need the stack.
                drop(stack);
                let value = Box::new((self.create)());
                return self.guard_stack(value);
            }
            // We're only here if we could get access to our stack, so just
            // create a new value. This seems like it could be wasteful, but
            // waiting for exclusive access to a stack when there's high
            // contention is brutal for perf.
            self.guard_stack_transient(Box::new((self.create)()))
        }

        /// Puts a value back into the pool. Callers don't need to call this.
        /// Once the guard that's returned by 'get' is dropped, it is put back
        /// into the pool automatically.
        #[inline]
        fn put_value(&self, value: Box<T>) {
            let caller = THREAD_ID.with(|id| *id);
            let stack_id = caller % self.stacks.len();
            // As with trying to pop a value from this thread's stack, we
            // merely attempt to get access to push this value back on the
            // stack. If there's too much contention, we just give up and throw
            // the value away.
            //
            // Interestingly, in ad hoc benchmarking, it is beneficial to
            // attempt to push the value back more than once, unlike when
            // popping the value. I don't have a good theory for why this is.
            // I guess if we drop too many values then that winds up forcing
            // the pop operation to create new fresh values and thus leads to
            // less reuse. There's definitely a balancing act here.
            for _ in 0..10 {
                let mut stack = match self.stacks[stack_id].0.try_lock() {
                    Err(_) => continue,
                    Ok(stack) => stack,
                };
                stack.push(value);
                return;
            }
        }

        /// Create a guard that represents the special owned T.
        #[inline]
        fn guard_owned(&self, caller: usize) -> PoolGuard<'_, T, F> {
            PoolGuard { pool: self, value: Err(caller), discard: false }
        }

        /// Create a guard that contains a value from the pool's stack.
        #[inline]
        fn guard_stack(&self, value: Box<T>) -> PoolGuard<'_, T, F> {
            PoolGuard { pool: self, value: Ok(value), discard: false }
        }

        /// Create a guard that contains a value from the pool's stack with an
        /// instruction to throw away the value instead of putting it back
        /// into the pool.
        #[inline]
        fn guard_stack_transient(&self, value: Box<T>) -> PoolGuard<'_, T, F> {
            PoolGuard { pool: self, value: Ok(value), discard: true }
        }
    }

    impl<T: core::fmt::Debug, F> core::fmt::Debug for Pool<T, F> {
        fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
            f.debug_struct("Pool")
                .field("stacks", &self.stacks)
                .field("owner", &self.owner)
                .field("owner_val", &self.owner_val)
                .finish()
        }
    }

    /// A guard that is returned when a caller requests a value from the pool.
    pub(super) struct PoolGuard<'a, T: Send, F: Fn() -> T> {
        /// The pool that this guard is attached to.
        pool: &'a Pool<T, F>,
        /// This is Err when the guard represents the special "owned" value.
        /// In which case, the value is retrieved from 'pool.owner_val'. And
        /// in the special case of `Err(THREAD_ID_DROPPED)`, it means the
        /// guard has been put back into the pool and should no longer be used.
        value: Result<Box<T>, usize>,
        /// When true, the value should be discarded instead of being pushed
        /// back into the pool. We tend to use this under high contention, and
        /// this allows us to avoid inflating the size of the pool. (Because
        /// under contention, we tend to create more values instead of waiting
        /// for access to a stack of existing values.)
        discard: bool,
    }

    impl<'a, T: Send, F: Fn() -> T> PoolGuard<'a, T, F> {
        /// Return the underlying value.
        #[inline]
        pub(super) fn value(&self) -> &T {
            match self.value {
                Ok(ref v) => &**v,
                // SAFETY: This is safe because the only way a PoolGuard gets
                // created for self.value=Err is when the current thread
                // corresponds to the owning thread, of which there can only
                // be one. Thus, we are guaranteed to be providing exclusive
                // access here which makes this safe.
                //
                // Also, since 'owner_val' is guaranteed to be initialized
                // before an owned PoolGuard is created, the unchecked unwrap
                // is safe.
                Err(id) => unsafe {
                    // This assert is *not* necessary for safety, since we
                    // should never be here if the guard had been put back into
                    // the pool. This is a sanity check to make sure we didn't
                    // break an internal invariant.
                    debug_assert_ne!(THREAD_ID_DROPPED, id);
                    (*self.pool.owner_val.get()).as_ref().unwrap_unchecked()
                },
            }
        }

        /// Return the underlying value as a mutable borrow.
        #[inline]
        pub(super) fn value_mut(&mut self) -> &mut T {
            match self.value {
                Ok(ref mut v) => &mut **v,
                // SAFETY: This is safe because the only way a PoolGuard gets
                // created for self.value=None is when the current thread
                // corresponds to the owning thread, of which there can only
                // be one. Thus, we are guaranteed to be providing exclusive
                // access here which makes this safe.
                //
                // Also, since 'owner_val' is guaranteed to be initialized
                // before an owned PoolGuard is created, the unwrap_unchecked
                // is safe.
                Err(id) => unsafe {
                    // This assert is *not* necessary for safety, since we
                    // should never be here if the guard had been put back into
                    // the pool. This is a sanity check to make sure we didn't
                    // break an internal invariant.
                    debug_assert_ne!(THREAD_ID_DROPPED, id);
                    (*self.pool.owner_val.get()).as_mut().unwrap_unchecked()
                },
            }
        }

        /// Consumes this guard and puts it back into the pool.
        #[inline]
        pub(super) fn put(this: PoolGuard<'_, T, F>) {
            // Since this is effectively consuming the guard and putting the
            // value back into the pool, there's no reason to run its Drop
            // impl after doing this. I don't believe there is a correctness
            // problem with doing so, but there's definitely a perf problem
            // by redoing this work. So we avoid it.
            let mut this = core::mem::ManuallyDrop::new(this);
            this.put_imp();
        }

        /// Puts this guard back into the pool by only borrowing the guard as
        /// mutable. This should be called at most once.
        #[inline(always)]
        fn put_imp(&mut self) {
            match core::mem::replace(&mut self.value, Err(THREAD_ID_DROPPED)) {
                Ok(value) => {
                    // If we were told to discard this value then don't bother
                    // trying to put it back into the pool. This occurs when
                    // the pop operation failed to acquire a lock and we
                    // decided to create a new value in lieu of contending for
                    // the lock.
                    if self.discard {
                        return;
                    }
                    self.pool.put_value(value);
                }
                // If this guard has a value "owned" by the thread, then
                // the Pool guarantees that this is the ONLY such guard.
                // Therefore, in order to place it back into the pool and make
                // it available, we need to change the owner back to the owning
                // thread's ID. But note that we use the ID that was stored in
                // the guard, since a guard can be moved to another thread and
                // dropped. (A previous iteration of this code read from the
                // THREAD_ID thread local, which uses the ID of the current
                // thread which may not be the ID of the owning thread! This
                // also avoids the TLS access, which is likely a hair faster.)
                Err(owner) => {
                    // If we hit this point, it implies 'put_imp' has been
                    // called multiple times for the same guard which in turn
                    // corresponds to a bug in this implementation.
                    assert_ne!(THREAD_ID_DROPPED, owner);
                    self.pool.owner.store(owner, Ordering::Release);
                }
            }
        }
    }

    impl<'a, T: Send, F: Fn() -> T> Drop for PoolGuard<'a, T, F> {
        #[inline]
        fn drop(&mut self) {
            self.put_imp();
        }
    }

    impl<'a, T: Send + core::fmt::Debug, F: Fn() -> T> core::fmt::Debug
        for PoolGuard<'a, T, F>
    {
        fn fmt(&self, f: &mut core::fmt::Formatter) -> core::fmt::Result {
            f.debug_struct("PoolGuard")
                .field("pool", &self.pool)
                .field("value", &self.value)
                .finish()
        }
    }
}

// FUTURE: We should consider using Mara Bos's nearly-lock-free version of this
// here: https://gist.github.com/m-ou-se/5fdcbdf7dcf4585199ce2de697f367a4.
//
// One reason why I did things with a "mutex" below is that it isolates the
// safety concerns to just the Mutex, where as the safety of Mara's pool is a
// bit more sprawling. I also expect this code to not be used that much, and
// so is unlikely to get as much real world usage with which to test it. That
// means the "obviously correct" lever is an important one.
//
// The specific reason to use Mara's pool is that it is likely faster and also
// less likely to hit problems with spin-locks, although it is not completely
// impervious to them.
//
// The best solution to this problem, probably, is a truly lock free pool. That
// could be done with a lock free linked list. The issue is the ABA problem. It
// is difficult to avoid, and doing so is complex. BUT, the upshot of that is
// that if we had a truly lock free pool, then we could also use it above in
// the 'std' pool instead of a Mutex because it should be completely free the
// problems that come from spin-locks.
#[cfg(not(feature = "std"))]
mod inner {
    use core::{
        cell::UnsafeCell,
        panic::{RefUnwindSafe, UnwindSafe},
        sync::atomic::{AtomicBool, Ordering},
    };

    use alloc::{boxed::Box, vec, vec::Vec};

    /// A thread safe pool utilizing alloc-only features.
    ///
    /// Unlike the std version, it doesn't seem possible(?) to implement the
    /// "thread owner" optimization because alloc-only doesn't have any concept
    /// of threads. So the best we can do is just a normal stack. This will
    /// increase latency in alloc-only environments.
    pub(super) struct Pool<T, F> {
        /// A stack of T values to hand out. These are used when a Pool is
        /// accessed by a thread that didn't create it.
        stack: Mutex<Vec<Box<T>>>,
        /// A function to create more T values when stack is empty and a caller
        /// has requested a T.
        create: F,
    }

    // If T is UnwindSafe, then since we provide exclusive access to any
    // particular value in the pool, it should therefore also be considered
    // RefUnwindSafe.
    impl<T: UnwindSafe, F: UnwindSafe> RefUnwindSafe for Pool<T, F> {}

    impl<T, F> Pool<T, F> {
        /// Create a new pool. The given closure is used to create values in
        /// the pool when necessary.
        pub(super) const fn new(create: F) -> Pool<T, F> {
            Pool { stack: Mutex::new(vec![]), create }
        }
    }

    impl<T: Send, F: Fn() -> T> Pool<T, F> {
        /// Get a value from the pool. This may block if another thread is also
        /// attempting to retrieve a value from the pool.
        #[inline]
        pub(super) fn get(&self) -> PoolGuard<'_, T, F> {
            let mut stack = self.stack.lock();
            let value = match stack.pop() {
                None => Box::new((self.create)()),
                Some(value) => value,
            };
            PoolGuard { pool: self, value: Some(value) }
        }

        #[inline]
        fn put(&self, guard: PoolGuard<'_, T, F>) {
            let mut guard = core::mem::ManuallyDrop::new(guard);
            if let Some(value) = guard.value.take() {
                self.put_value(value);
            }
        }

        /// Puts a value back into the pool. Callers don't need to call this.
        /// Once the guard that's returned by 'get' is dropped, it is put back
        /// into the pool automatically.
        #[inline]
        fn put_value(&self, value: Box<T>) {
            let mut stack = self.stack.lock();
            stack.push(value);
        }
    }

    impl<T: core::fmt::Debug, F> core::fmt::Debug for Pool<T, F> {
        fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
            f.debug_struct("Pool").field("stack", &self.stack).finish()
        }
    }

    /// A guard that is returned when a caller requests a value from the pool.
    pub(super) struct PoolGuard<'a, T: Send, F: Fn() -> T> {
        /// The pool that this guard is attached to.
        pool: &'a Pool<T, F>,
        /// This is None after the guard has been put back into the pool.
        value: Option<Box<T>>,
    }

    impl<'a, T: Send, F: Fn() -> T> PoolGuard<'a, T, F> {
        /// Return the underlying value.
        #[inline]
        pub(super) fn value(&self) -> &T {
            self.value.as_deref().unwrap()
        }

        /// Return the underlying value as a mutable borrow.
        #[inline]
        pub(super) fn value_mut(&mut self) -> &mut T {
            self.value.as_deref_mut().unwrap()
        }

        /// Consumes this guard and puts it back into the pool.
        #[inline]
        pub(super) fn put(this: PoolGuard<'_, T, F>) {
            // Since this is effectively consuming the guard and putting the
            // value back into the pool, there's no reason to run its Drop
            // impl after doing this. I don't believe there is a correctness
            // problem with doing so, but there's definitely a perf problem
            // by redoing this work. So we avoid it.
            let mut this = core::mem::ManuallyDrop::new(this);
            this.put_imp();
        }

        /// Puts this guard back into the pool by only borrowing the guard as
        /// mutable. This should be called at most once.
        #[inline(always)]
        fn put_imp(&mut self) {
            if let Some(value) = self.value.take() {
                self.pool.put_value(value);
            }
        }
    }

    impl<'a, T: Send, F: Fn() -> T> Drop for PoolGuard<'a, T, F> {
        #[inline]
        fn drop(&mut self) {
            self.put_imp();
        }
    }

    impl<'a, T: Send + core::fmt::Debug, F: Fn() -> T> core::fmt::Debug
        for PoolGuard<'a, T, F>
    {
        fn fmt(&self, f: &mut core::fmt::Formatter) -> core::fmt::Result {
            f.debug_struct("PoolGuard")
                .field("pool", &self.pool)
                .field("value", &self.value)
                .finish()
        }
    }

    /// A spin-lock based mutex. Yes, I have read spinlocks cosnidered
    /// harmful[1], and if there's a reasonable alternative choice, I'll
    /// happily take it.
    ///
    /// I suspect the most likely alternative here is a Treiber stack, but
    /// implementing one correctly in a way that avoids the ABA problem looks
    /// subtle enough that I'm not sure I want to attempt that. But otherwise,
    /// we only need a mutex in order to implement our pool, so if there's
    /// something simpler we can use that works for our `Pool` use case, then
    /// that would be great.
    ///
    /// Note that this mutex does not do poisoning.
    ///
    /// [1]: https://matklad.github.io/2020/01/02/spinlocks-considered-harmful.html
    #[derive(Debug)]
    struct Mutex<T> {
        locked: AtomicBool,
        data: UnsafeCell<T>,
    }

    // SAFETY: Since a Mutex guarantees exclusive access, as long as we can
    // send it across threads, it must also be Sync.
    unsafe impl<T: Send> Sync for Mutex<T> {}

    impl<T> Mutex<T> {
        /// Create a new mutex for protecting access to the given value across
        /// multiple threads simultaneously.
        const fn new(value: T) -> Mutex<T> {
            Mutex {
                locked: AtomicBool::new(false),
                data: UnsafeCell::new(value),
            }
        }

        /// Lock this mutex and return a guard providing exclusive access to
        /// `T`. This blocks if some other thread has already locked this
        /// mutex.
        #[inline]
        fn lock(&self) -> MutexGuard<'_, T> {
            while self
                .locked
                .compare_exchange(
                    false,
                    true,
                    Ordering::AcqRel,
                    Ordering::Acquire,
                )
                .is_err()
            {
                core::hint::spin_loop();
            }
            // SAFETY: The only way we're here is if we successfully set
            // 'locked' to true, which implies we must be the only thread here
            // and thus have exclusive access to 'data'.
            let data = unsafe { &mut *self.data.get() };
            MutexGuard { locked: &self.locked, data }
        }
    }

    /// A guard that derefs to &T and &mut T. When it's dropped, the lock is
    /// released.
    #[derive(Debug)]
    struct MutexGuard<'a, T> {
        locked: &'a AtomicBool,
        data: &'a mut T,
    }

    impl<'a, T> core::ops::Deref for MutexGuard<'a, T> {
        type Target = T;

        #[inline]
        fn deref(&self) -> &T {
            self.data
        }
    }

    impl<'a, T> core::ops::DerefMut for MutexGuard<'a, T> {
        #[inline]
        fn deref_mut(&mut self) -> &mut T {
            self.data
        }
    }

    impl<'a, T> Drop for MutexGuard<'a, T> {
        #[inline]
        fn drop(&mut self) {
            // Drop means 'data' is no longer accessible, so we can unlock
            // the mutex.
            self.locked.store(false, Ordering::Release);
        }
    }
}

#[cfg(test)]
mod tests {
    use core::panic::{RefUnwindSafe, UnwindSafe};

    use alloc::{boxed::Box, vec, vec::Vec};

    use super::*;

    #[test]
    fn oibits() {
        fn assert_oitbits<T: Send + Sync + UnwindSafe + RefUnwindSafe>() {}
        assert_oitbits::<Pool<Vec<u32>>>();
        assert_oitbits::<Pool<core::cell::RefCell<Vec<u32>>>>();
        assert_oitbits::<
            Pool<
                Vec<u32>,
                Box<
                    dyn Fn() -> Vec<u32>
                        + Send
                        + Sync
                        + UnwindSafe
                        + RefUnwindSafe,
                >,
            >,
        >();
    }

    // Tests that Pool implements the "single owner" optimization. That is, the
    // thread that first accesses the pool gets its own copy, while all other
    // threads get distinct copies.
    #[cfg(feature = "std")]
    #[test]
    fn thread_owner_optimization() {
        use std::{cell::RefCell, sync::Arc, vec};

        let pool: Arc<Pool<RefCell<Vec<char>>>> =
            Arc::new(Pool::new(|| RefCell::new(vec!['a'])));
        pool.get().borrow_mut().push('x');

        let pool1 = pool.clone();
        let t1 = std::thread::spawn(move || {
            let guard = pool1.get();
            guard.borrow_mut().push('y');
        });

        let pool2 = pool.clone();
        let t2 = std::thread::spawn(move || {
            let guard = pool2.get();
            guard.borrow_mut().push('z');
        });

        t1.join().unwrap();
        t2.join().unwrap();

        // If we didn't implement the single owner optimization, then one of
        // the threads above is likely to have mutated the [a, x] vec that
        // we stuffed in the pool before spawning the threads. But since
        // neither thread was first to access the pool, and because of the
        // optimization, we should be guaranteed that neither thread mutates
        // the special owned pool value.
        //
        // (Technically this is an implementation detail and not a contract of
        // Pool's API.)
        assert_eq!(vec!['a', 'x'], *pool.get().borrow());
    }

    // This tests that if the "owner" of a pool asks for two values, then it
    // gets two distinct values and not the same one. This test failed in the
    // course of developing the pool, which in turn resulted in UB because it
    // permitted getting aliasing &mut borrows to the same place in memory.
    #[test]
    fn thread_owner_distinct() {
        let pool = Pool::new(|| vec!['a']);

        {
            let mut g1 = pool.get();
            let v1 = &mut *g1;
            let mut g2 = pool.get();
            let v2 = &mut *g2;
            v1.push('b');
            v2.push('c');
            assert_eq!(&mut vec!['a', 'b'], v1);
            assert_eq!(&mut vec!['a', 'c'], v2);
        }
        // This isn't technically guaranteed, but we
        // expect to now get the "owned" value (the first
        // call to 'get()' above) now that it's back in
        // the pool.
        assert_eq!(&mut vec!['a', 'b'], &mut *pool.get());
    }

    // This tests that we can share a guard with another thread, mutate the
    // underlying value and everything works. This failed in the course of
    // developing a pool since the pool permitted 'get()' to return the same
    // value to the owner thread, even before the previous value was put back
    // into the pool. This in turn resulted in this test producing a data race.
    #[cfg(feature = "std")]
    #[test]
    fn thread_owner_sync() {
        let pool = Pool::new(|| vec!['a']);
        {
            let mut g1 = pool.get();
            let mut g2 = pool.get();
            std::thread::scope(|s| {
                s.spawn(|| {
                    g1.push('b');
                });
                s.spawn(|| {
                    g2.push('c');
                });
            });

            let v1 = &mut *g1;
            let v2 = &mut *g2;
            assert_eq!(&mut vec!['a', 'b'], v1);
            assert_eq!(&mut vec!['a', 'c'], v2);
        }

        // This isn't technically guaranteed, but we
        // expect to now get the "owned" value (the first
        // call to 'get()' above) now that it's back in
        // the pool.
        assert_eq!(&mut vec!['a', 'b'], &mut *pool.get());
    }

    // This tests that if we move a PoolGuard that is owned by the current
    // thread to another thread and drop it, then the thread owner doesn't
    // change. During development of the pool, this test failed because the
    // PoolGuard assumed it was dropped in the same thread from which it was
    // created, and thus used the current thread's ID as the owner, which could
    // be different than the actual owner of the pool.
    #[cfg(feature = "std")]
    #[test]
    fn thread_owner_send_drop() {
        let pool = Pool::new(|| vec!['a']);
        // Establishes this thread as the owner.
        {
            pool.get().push('b');
        }
        std::thread::scope(|s| {
            // Sanity check that we get the same value back.
            // (Not technically guaranteed.)
            let mut g = pool.get();
            assert_eq!(&vec!['a', 'b'], &*g);
            // Now push it to another thread and drop it.
            s.spawn(move || {
                g.push('c');
            })
            .join()
            .unwrap();
        });
        // Now check that we're still the owner. This is not technically
        // guaranteed by the API, but is true in practice given the thread
        // owner optimization.
        assert_eq!(&vec!['a', 'b', 'c'], &*pool.get());
    }
}