mmtk/util/alloc/

allocator.rs

use crate::global_state::GlobalState;
use crate::util::address::Address;
#[cfg(feature = "analysis")]
use crate::util::analysis::AnalysisManager;
use crate::util::heap::gc_trigger::GCTrigger;
use crate::util::options::Options;
use crate::MMTK;

use std::cell::RefCell;
use std::sync::atomic::Ordering;
use std::sync::Arc;

use crate::policy::space::Space;
use crate::util::opaque_pointer::*;
use crate::vm::VMBinding;
use crate::vm::{ActivePlan, Collection};
use downcast_rs::Downcast;

#[repr(C)]
#[derive(Debug)]
/// A list of errors that MMTk can encounter during allocation.
pub enum AllocationError {
    /// The specified heap size is too small for the given program to continue.
    HeapOutOfMemory,
    /// The OS is unable to mmap or acquire more memory. Critical error. MMTk expects the VM to
    /// abort if such an error is thrown.
    MmapOutOfMemory,
}

/// Allow specifying different behaviors with [`Allocator::alloc_with_options`].
#[repr(C)]
#[derive(Copy, Clone, PartialEq, Eq, Debug)]
pub struct AllocationOptions {
    /// Whether over-committing is allowed at this allocation site.  Over-committing means the
    /// allocation is allowed to go beyond the current heap size.  But it is not guaranteed to
    /// succeed.
    ///
    /// **The default is `false`**.
    ///
    /// Note that regardless of the value of `allow_overcommit`, the allocation may trigger GC if
    /// the GC trigger considers it needed.
    pub allow_overcommit: bool,

    /// Whether the allocation is at a safepoint.
    ///
    /// **The default is `true`**.
    ///
    /// If `true`, the allocation is allowed to block for GC.
    ///
    /// If `false`, the allocation will immediately return a null address if the allocation cannot
    /// be satisfied without a GC.
    pub at_safepoint: bool,

    /// Whether the allocation is allowed to call [`Collection::out_of_memory`].
    ///
    /// **The default is `true`**.
    ///
    /// If `true`, the allocation will call [`Collection::out_of_memory`] when out of memory and
    /// return null.
    ///
    /// If `fasle`, the allocation will return null immediately when out of memory.
    pub allow_oom_call: bool,
}

/// The default value for `AllocationOptions` has the same semantics as calling [`Allocator::alloc`]
/// directly.
impl Default for AllocationOptions {
    fn default() -> Self {
        Self {
            allow_overcommit: false,
            at_safepoint: true,
            allow_oom_call: true,
        }
    }
}

impl AllocationOptions {
    pub(crate) fn is_default(&self) -> bool {
        *self == AllocationOptions::default()
    }
}

/// A wrapper for [`AllocatorContext`] to hold a [`AllocationOptions`] that can be modified by the
/// same mutator thread.
///
/// All [`Allocator`] instances in `Allocators` share one `AllocationOptions` instance, and it will
/// only be accessed by the mutator (via `Mutator::allocators`) or the GC worker (via
/// `GCWorker::copy`) that owns it.  Rust doesn't like multiple mutable references pointing to a
/// shared data structure.  We cannot use [`atomic::Atomic`] because `AllocationOptions` has
/// multiple fields. We wrap it in a `RefCell` to make it internally mutable.
///
/// Note: The allocation option is called every time [`Allocator::alloc_with_options`] is called.
/// Because API functions should only be called on allocation slow paths, we believe that `RefCell`
/// should be good enough for performance.  If this is too slow, we may consider `UnsafeCell`.  If
/// that's still too slow, we should consider changing the API to make the allocation options a
/// persistent per-mutator value, and allow the VM binding set its value via a new API function.
struct AllocationOptionsHolder {
    alloc_options: RefCell<AllocationOptions>,
}

/// Strictly speaking, `AllocationOptionsHolder` isn't `Sync`.  Two threads cannot set or clear the
/// same `AllocationOptionsHolder` at the same time.  However, both `Mutator` and `GCWorker` are
/// `Send`, and both of which own `Allocators` and require its field `Arc<AllocationContext>` to be
/// `Send`, which requires `AllocationContext` to be `Sync`, which requires
/// `AllocationOptionsHolder` to be `Sync`.  (Note that `Arc<T>` can be cloned and given to another
/// thread, and Rust expects `T` to be `Sync`, too.  But we never share `AllocationContext` between
/// threads, but only between multiple `Allocator` instances within the same `Allocators` instance.
/// Rust can't figure this out.)
unsafe impl Sync for AllocationOptionsHolder {}

impl AllocationOptionsHolder {
    pub fn new(alloc_options: AllocationOptions) -> Self {
        Self {
            alloc_options: RefCell::new(alloc_options),
        }
    }
    pub fn set_alloc_options(&self, options: AllocationOptions) {
        let mut alloc_options = self.alloc_options.borrow_mut();
        *alloc_options = options;
    }

    pub fn clear_alloc_options(&self) {
        let mut alloc_options = self.alloc_options.borrow_mut();
        *alloc_options = AllocationOptions::default();
    }

    pub fn get_alloc_options(&self) -> AllocationOptions {
        let alloc_options = self.alloc_options.borrow();
        *alloc_options
    }
}

pub fn align_allocation_no_fill<VM: VMBinding>(
    region: Address,
    alignment: usize,
    offset: usize,
) -> Address {
    align_allocation_inner::<VM>(region, alignment, offset, VM::MIN_ALIGNMENT, false)
}

pub fn align_allocation<VM: VMBinding>(
    region: Address,
    alignment: usize,
    offset: usize,
) -> Address {
    align_allocation_inner::<VM>(region, alignment, offset, VM::MIN_ALIGNMENT, true)
}

pub fn align_allocation_inner<VM: VMBinding>(
    region: Address,
    alignment: usize,
    offset: usize,
    known_alignment: usize,
    fillalignmentgap: bool,
) -> Address {
    debug_assert!(known_alignment >= VM::MIN_ALIGNMENT);
    // Make sure MIN_ALIGNMENT is reasonable.
    #[allow(clippy::assertions_on_constants)]
    {
        // TODO: This is a static assertion that VM::MIN_ALIGNMENT must be at least 4.
        // This assertion has existed since JikesRVM MMTk.
        // We are keeping it here because some implementation details of the allocator may rely on this assertion.
        // Some GC algorithms may require a stricter minimum alignment, and that can override the value.
        // We should refactor the VM binding API and the internal interface
        // to reconcile the requirements from the VM and the GC algorithms.
        debug_assert!(VM::MIN_ALIGNMENT >= std::mem::size_of::<i32>());
    }
    debug_assert!(!(fillalignmentgap && region.is_zero()));
    debug_assert!(alignment <= VM::MAX_ALIGNMENT);
    debug_assert!(region.is_aligned_to(VM::ALLOC_END_ALIGNMENT));
    debug_assert!((alignment & (VM::MIN_ALIGNMENT - 1)) == 0);
    debug_assert!((offset & (VM::MIN_ALIGNMENT - 1)) == 0);

    // No alignment ever required.
    if alignment <= known_alignment || VM::MAX_ALIGNMENT <= VM::MIN_ALIGNMENT {
        return region;
    }

    // May require an alignment
    let mask = (alignment - 1) as isize; // fromIntSignExtend
    let neg_off: isize = -(offset as isize); // fromIntSignExtend
    let delta = neg_off.wrapping_sub_unsigned(region.as_usize()) & mask; // Use wrapping_sub to avoid overflow

    if fillalignmentgap && (VM::ALIGNMENT_VALUE != 0) {
        fill_alignment_gap::<VM>(region, region + delta);
    }

    region + delta
}

/// Fill the specified region with the alignment value.
pub fn fill_alignment_gap<VM: VMBinding>(start: Address, end: Address) {
    if VM::ALIGNMENT_VALUE != 0 {
        let start_ptr = start.to_mut_ptr::<u8>();
        unsafe {
            std::ptr::write_bytes(start_ptr, VM::ALIGNMENT_VALUE, end - start);
        }
    }
}

pub fn get_maximum_aligned_size<VM: VMBinding>(size: usize, alignment: usize) -> usize {
    get_maximum_aligned_size_inner::<VM>(size, alignment, VM::MIN_ALIGNMENT)
}

pub fn get_maximum_aligned_size_inner<VM: VMBinding>(
    size: usize,
    alignment: usize,
    known_alignment: usize,
) -> usize {
    trace!(
        "size={}, alignment={}, known_alignment={}, MIN_ALIGNMENT={}",
        size,
        alignment,
        known_alignment,
        VM::MIN_ALIGNMENT
    );
    debug_assert!(size == size & !(known_alignment - 1));
    debug_assert!(known_alignment >= VM::MIN_ALIGNMENT);

    if VM::MAX_ALIGNMENT <= VM::MIN_ALIGNMENT || alignment <= known_alignment {
        size
    } else {
        size + alignment - known_alignment
    }
}

#[cfg(debug_assertions)]
pub(crate) fn assert_allocation_args<VM: VMBinding>(size: usize, align: usize, offset: usize) {
    use crate::util::constants::*;
    // MMTk has assumptions about minimal object size.
    // We need to make sure that all allocations comply with the min object size.
    // Ideally, we check the allocation size, and if it is smaller, we transparently allocate the min
    // object size (the VM does not need to know this). However, for the VM bindings we support at the moment,
    // their object sizes are all larger than MMTk's min object size, so we simply put an assertion here.
    // If you plan to use MMTk with a VM with its object size smaller than MMTk's min object size, you should
    // meet the min object size in the fastpath.
    debug_assert!(size >= MIN_OBJECT_SIZE);
    // Assert alignment
    debug_assert!(align >= VM::MIN_ALIGNMENT);
    debug_assert!(align <= VM::MAX_ALIGNMENT);
    // Assert offset
    debug_assert!(VM::USE_ALLOCATION_OFFSET || offset == 0);
}

/// The context an allocator needs to access in order to perform allocation.
pub struct AllocatorContext<VM: VMBinding> {
    alloc_options: AllocationOptionsHolder,
    pub state: Arc<GlobalState>,
    pub options: Arc<Options>,
    pub gc_trigger: Arc<GCTrigger<VM>>,
    #[cfg(feature = "analysis")]
    pub analysis_manager: Arc<AnalysisManager<VM>>,
}

impl<VM: VMBinding> AllocatorContext<VM> {
    pub fn new(mmtk: &MMTK<VM>) -> Self {
        Self {
            alloc_options: AllocationOptionsHolder::new(AllocationOptions::default()),
            state: mmtk.state.clone(),
            options: mmtk.options.clone(),
            gc_trigger: mmtk.gc_trigger.clone(),
            #[cfg(feature = "analysis")]
            analysis_manager: mmtk.analysis_manager.clone(),
        }
    }

    pub fn set_alloc_options(&self, options: AllocationOptions) {
        self.alloc_options.set_alloc_options(options);
    }

    pub fn clear_alloc_options(&self) {
        self.alloc_options.clear_alloc_options();
    }

    pub fn get_alloc_options(&self) -> AllocationOptions {
        self.alloc_options.get_alloc_options()
    }
}

/// A trait which implements allocation routines. Every allocator needs to implements this trait.
pub trait Allocator<VM: VMBinding>: Downcast {
    /// Return the [`VMThread`] associated with this allocator instance.
    fn get_tls(&self) -> VMThread;

    /// Return the [`Space`](src/policy/space/Space) instance associated with this allocator instance.
    fn get_space(&self) -> &'static dyn Space<VM>;

    /// Return the context for the allocator.
    fn get_context(&self) -> &AllocatorContext<VM>;

    /// Return if this allocator can do thread local allocation. If an allocator does not do thread
    /// local allocation, each allocation will go to slowpath and will have a check for GC polls.
    fn does_thread_local_allocation(&self) -> bool;

    /// Return at which granularity the allocator acquires memory from the global space and use
    /// them as thread local buffer. For example, the [`BumpAllocator`](crate::util::alloc::BumpAllocator) acquires memory at 32KB
    /// blocks. Depending on the actual size for the current object, they always acquire memory of
    /// N*32KB (N>=1). Thus the [`BumpAllocator`](crate::util::alloc::BumpAllocator) returns 32KB for this method.  Only allocators
    /// that do thread local allocation need to implement this method.
    fn get_thread_local_buffer_granularity(&self) -> usize {
        assert!(self.does_thread_local_allocation(), "An allocator that does not thread local allocation does not have a buffer granularity.");
        unimplemented!()
    }

    /// An allocation attempt. The implementation of this function depends on the allocator used.
    /// If an allocator supports thread local allocations, then the allocation will be serviced
    /// from its TLAB, otherwise it will default to using the slowpath, i.e. [`alloc_slow`](Allocator::alloc_slow).
    ///
    /// If the heap is full, we trigger a GC and attempt to free up
    /// more memory, and re-attempt the allocation.
    ///
    /// Note that in the case where the VM is out of memory, we invoke
    /// [`Collection::out_of_memory`] to inform the binding and then return a null pointer back to
    /// it. We have no assumptions on whether the VM will continue executing or abort immediately.
    /// If the VM continues execution, the function will return a null address.
    ///
    /// An allocator needs to make sure the object reference for the returned address is in the same
    /// chunk as the returned address (so the side metadata and the SFT for an object reference is valid).
    /// See [`crate::util::alloc::object_ref_guard`](util/alloc/object_ref_guard).
    ///
    /// Arguments:
    /// * `size`: the allocation size in bytes.
    /// * `align`: the required alignment in bytes.
    /// * `offset` the required offset in bytes.
    fn alloc(&mut self, size: usize, align: usize, offset: usize) -> Address;

    /// An allocation attempt. The allocation options may specify different behaviors for this allocation request.
    ///
    /// Arguments:
    /// * `size`: the allocation size in bytes.
    /// * `align`: the required alignment in bytes.
    /// * `offset` the required offset in bytes.
    /// * `options`: the allocation options to change the default allocation behavior for this request.
    fn alloc_with_options(
        &mut self,
        size: usize,
        align: usize,
        offset: usize,
        alloc_options: AllocationOptions,
    ) -> Address {
        self.get_context().set_alloc_options(alloc_options);
        let ret = self.alloc(size, align, offset);
        self.get_context().clear_alloc_options();
        ret
    }

    /// Slowpath allocation attempt. This function is explicitly not inlined for performance
    /// considerations.
    ///
    /// Arguments:
    /// * `size`: the allocation size in bytes.
    /// * `align`: the required alignment in bytes.
    /// * `offset` the required offset in bytes.
    #[inline(never)]
    fn alloc_slow(&mut self, size: usize, align: usize, offset: usize) -> Address {
        self.alloc_slow_inline(size, align, offset)
    }

    /// Slowpath allocation attempt. Mostly the same as [`Allocator::alloc_slow`], except that the allocation options
    /// may specify different behaviors for this allocation request.
    ///
    /// This function is not used internally. It is mostly for the bindings.
    /// [`Allocator::alloc_with_options`] still calls the normal [`Allocator::alloc_slow`].
    ///
    /// Arguments:
    /// * `size`: the allocation size in bytes.
    /// * `align`: the required alignment in bytes.
    /// * `offset` the required offset in bytes.
    fn alloc_slow_with_options(
        &mut self,
        size: usize,
        align: usize,
        offset: usize,
        alloc_options: AllocationOptions,
    ) -> Address {
        // The function is not used internally. We won't set no_gc_on_fail redundantly.
        self.get_context().set_alloc_options(alloc_options);
        let ret = self.alloc_slow(size, align, offset);
        self.get_context().clear_alloc_options();
        ret
    }

    /// Slowpath allocation attempt. This function executes the actual slowpath allocation.  A
    /// slowpath allocation in MMTk attempts to allocate the object using the per-allocator
    /// definition of [`alloc_slow_once`](Allocator::alloc_slow_once). This function also accounts for increasing the
    /// allocation bytes in order to support stress testing. In case precise stress testing is
    /// being used, the [`alloc_slow_once_precise_stress`](Allocator::alloc_slow_once_precise_stress) function is used instead.
    ///
    /// Note that in the case where the VM is out of memory, we invoke
    /// [`Collection::out_of_memory`] with a [`AllocationError::HeapOutOfMemory`] error to inform
    /// the binding and then return a null pointer back to it. We have no assumptions on whether
    /// the VM will continue executing or abort immediately on a
    /// [`AllocationError::HeapOutOfMemory`] error.
    ///
    /// Arguments:
    /// * `size`: the allocation size in bytes.
    /// * `align`: the required alignment in bytes.
    /// * `offset` the required offset in bytes.
    fn alloc_slow_inline(&mut self, size: usize, align: usize, offset: usize) -> Address {
        let tls = self.get_tls();
        let is_mutator = VM::VMActivePlan::is_mutator(tls);
        let stress_test = self.get_context().options.is_stress_test_gc_enabled();

        // Information about the previous collection.
        let mut emergency_collection = false;
        let mut previous_result_zero = false;

        loop {
            // Try to allocate using the slow path
            let result = if is_mutator && stress_test && *self.get_context().options.precise_stress
            {
                // If we are doing precise stress GC, we invoke the special allow_slow_once call.
                // alloc_slow_once_precise_stress() should make sure that every allocation goes
                // to the slowpath (here) so we can check the allocation bytes and decide
                // if we need to do a stress GC.

                // If we should do a stress GC now, we tell the alloc_slow_once_precise_stress()
                // so they would avoid try any thread local allocation, and directly call
                // global acquire and do a poll.
                let need_poll = is_mutator && self.get_context().gc_trigger.should_do_stress_gc();
                self.alloc_slow_once_precise_stress(size, align, offset, need_poll)
            } else {
                // If we are not doing precise stress GC, just call the normal alloc_slow_once().
                // Normal stress test only checks for stress GC in the slowpath.
                self.alloc_slow_once_traced(size, align, offset)
            };

            if !is_mutator {
                debug_assert!(!result.is_zero());
                return result;
            }

            if !result.is_zero() {
                // Report allocation success to assist OutOfMemory handling.
                if !self
                    .get_context()
                    .state
                    .allocation_success
                    .load(Ordering::Relaxed)
                {
                    self.get_context()
                        .state
                        .allocation_success
                        .store(true, Ordering::SeqCst);
                }

                // Only update the allocation bytes if we haven't failed a previous allocation in this loop
                if stress_test && self.get_context().state.is_initialized() && !previous_result_zero
                {
                    let allocated_size = if *self.get_context().options.precise_stress
                        || !self.does_thread_local_allocation()
                    {
                        // For precise stress test, or for allocators that do not have thread local buffer,
                        // we know exactly how many bytes we allocate.
                        size
                    } else {
                        // For normal stress test, we count the entire thread local buffer size as allocated.
                        crate::util::conversions::raw_align_up(
                            size,
                            self.get_thread_local_buffer_granularity(),
                        )
                    };
                    let _allocation_bytes = self
                        .get_context()
                        .state
                        .increase_allocation_bytes_by(allocated_size);

                    // This is the allocation hook for the analysis trait. If you want to call
                    // an analysis counter specific allocation hook, then here is the place to do so
                    #[cfg(feature = "analysis")]
                    if _allocation_bytes > *self.get_context().options.analysis_factor {
                        trace!(
                            "Analysis: allocation_bytes = {} more than analysis_factor = {}",
                            _allocation_bytes,
                            *self.get_context().options.analysis_factor
                        );

                        self.get_context()
                            .analysis_manager
                            .alloc_hook(size, align, offset);
                    }
                }

                return result;
            }

            // From here on, we handle the case that alloc_once failed.
            assert!(result.is_zero());

            if !self.get_context().get_alloc_options().at_safepoint {
                // If the allocation is not at safepoint, it will not be able to block for GC.  But
                // the code beyond this point tests OOM conditions and, if not OOM, try to allocate
                // again.  Since we didn't block for GC, the allocation will fail again if we try
                // again. So we return null immediately.
                return Address::ZERO;
            }

            // It is possible to have cases where a thread is blocked for another GC (non emergency)
            // immediately after being blocked for a GC (emergency) (e.g. in stress test), that is saying
            // the thread does not leave this loop between the two GCs. The local var 'emergency_collection'
            // was set to true after the first GC. But when we execute this check below, we just finished
            // the second GC, which is not emergency. In such case, we will give a false OOM.
            // We cannot just rely on the local var. Instead, we get the emergency collection value again,
            // and check both.
            if emergency_collection && self.get_context().state.is_emergency_collection() {
                trace!("Emergency collection");
                // Report allocation success to assist OutOfMemory handling.
                // This seems odd, but we must allow each OOM to run its course (and maybe give us back memory)
                let fail_with_oom = !self
                    .get_context()
                    .state
                    .allocation_success
                    .swap(true, Ordering::SeqCst);
                trace!("fail with oom={}", fail_with_oom);
                if fail_with_oom {
                    // Note that we throw a `HeapOutOfMemory` error here and return a null ptr back to the VM
                    trace!("Throw HeapOutOfMemory!");
                    VM::VMCollection::out_of_memory(tls, AllocationError::HeapOutOfMemory);
                    self.get_context()
                        .state
                        .allocation_success
                        .store(false, Ordering::SeqCst);
                    return result;
                }
            }

            /* This is in case a GC occurs, and our mutator context is stale.
             * In some VMs the scheduler can change the affinity between the
             * current thread and the mutator context. This is possible for
             * VMs that dynamically multiplex Java threads onto multiple mutator
             * contexts. */
            // FIXME: No good way to do this
            //current = unsafe {
            //    VMActivePlan::mutator(tls).get_allocator_from_space(space)
            //};

            // Record whether last collection was an Emergency collection. If so, we make one more
            // attempt to allocate before we signal an OOM.
            emergency_collection = self.get_context().state.is_emergency_collection();
            trace!("Got emergency collection as {}", emergency_collection);
            previous_result_zero = true;
        }
    }

    /// Single slow path allocation attempt. This is called by [`alloc_slow_inline`](Allocator::alloc_slow_inline). The
    /// implementation of this function depends on the allocator used. Generally, if an allocator
    /// supports thread local allocations, it will try to allocate more TLAB space here. If it
    /// doesn't, then (generally) the allocator simply allocates enough space for the current
    /// object.
    ///
    /// Arguments:
    /// * `size`: the allocation size in bytes.
    /// * `align`: the required alignment in bytes.
    /// * `offset` the required offset in bytes.
    fn alloc_slow_once(&mut self, size: usize, align: usize, offset: usize) -> Address;

    /// A wrapper method for [`alloc_slow_once`](Allocator::alloc_slow_once) to insert USDT tracepoints.
    ///
    /// Arguments:
    /// * `size`: the allocation size in bytes.
    /// * `align`: the required alignment in bytes.
    /// * `offset` the required offset in bytes.
    fn alloc_slow_once_traced(&mut self, size: usize, align: usize, offset: usize) -> Address {
        probe!(mmtk, alloc_slow_once_start);
        // probe! expands to an empty block on unsupported platforms
        #[allow(clippy::let_and_return)]
        let ret = self.alloc_slow_once(size, align, offset);
        probe!(mmtk, alloc_slow_once_end);
        ret
    }

    /// Single slowpath allocation attempt for stress test. When the stress factor is set (e.g. to
    /// N), we would expect for every N bytes allocated, we will trigger a stress GC.  However, for
    /// allocators that do thread local allocation, they may allocate from their thread local
    /// buffer which does not have a GC poll check, and they may even allocate with the JIT
    /// generated allocation fastpath which is unaware of stress test GC. For both cases, we are
    /// not able to guarantee a stress GC is triggered every N bytes. To solve this, when the
    /// stress factor is set, we will call this method instead of the normal alloc_slow_once(). We
    /// expect the implementation of this slow allocation will trick the fastpath so every
    /// allocation will fail in the fastpath, jump to the slow path and eventually call this method
    /// again for the actual allocation.
    ///
    /// The actual implementation about how to trick the fastpath may vary. For example, our bump
    /// pointer allocator will set the thread local buffer limit to the buffer size instead of the
    /// buffer end address. In this case, every fastpath check (cursor + size < limit) will fail,
    /// and jump to this slowpath. In the slowpath, we still allocate from the thread local buffer,
    /// and recompute the limit (remaining buffer size).
    ///
    /// If an allocator does not do thread local allocation (which returns false for
    /// does_thread_local_allocation()), it does not need to override this method. The default
    /// implementation will simply call allow_slow_once() and it will work fine for allocators that
    /// do not have thread local allocation.
    ///
    /// Arguments:
    /// * `size`: the allocation size in bytes.
    /// * `align`: the required alignment in bytes.
    /// * `offset` the required offset in bytes.
    /// * `need_poll`: if this is true, the implementation must poll for a GC, rather than
    ///   attempting to allocate from the local buffer.
    fn alloc_slow_once_precise_stress(
        &mut self,
        size: usize,
        align: usize,
        offset: usize,
        need_poll: bool,
    ) -> Address {
        // If an allocator does thread local allocation but does not override this method to
        // provide a correct implementation, we will log a warning.
        if self.does_thread_local_allocation() && need_poll {
            warn!("{} does not support stress GC (An allocator that does thread local allocation needs to implement allow_slow_once_stress_test()).", std::any::type_name::<Self>());
        }
        self.alloc_slow_once_traced(size, align, offset)
    }

    /// The [`crate::plan::Mutator`] that includes this allocator is going to be destroyed. Some allocators
    /// may need to save/transfer its thread local data to the space.
    fn on_mutator_destroy(&mut self) {
        // By default, do nothing
    }
}

impl_downcast!(Allocator<VM> where VM: VMBinding);