Inline GC handle write barrier

[Copilot]

Summary

• Inline the GC handle write-barrier worker into its callers without changing the GC/EE interface.
• Avoid the handle-type lookup when the assigned target is in generation zero.
• Relates to [Perf] Linux/x64: 2 Regressions on 7/5/2026 4:58:39 PM +00:00 #130698.

Validation

• Built the CoreCLR GC native target successfully.
• Built libraries successfully.
• Full CoreCLR Release build reached the changed GC objects successfully but failed later on pre-existing JIT identifier errors in instrsxarch.h.
• The focused GCHandle.Target test could not be built because test dependency assets have not been generated.

Performance

• EgorBot measurement requested for Linux x64 and macOS ARM64.

Note

This pull request description was generated by AI/Copilot.

[azure-pipelines]

Azure Pipelines:
Successfully started running 3 pipeline(s).
13 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[dotnet-policy-service]

Tagging subscribers to this area: @anicka-net, @dotnet/gc
See info in area-owners.md if you want to be subscribed.

[Copilot]

Pull request overview

This PR refactors the CoreCLR GC handle write-barrier by moving the HndWriteBarrierWorker implementation into handletable.inl (so it can be inlined into its callers) and introduces a fast-path that avoids the handle-type lookup when the assigned target is in generation 0. This targets reduced overhead in hot handle assignment paths while keeping the GC/EE interface shape unchanged.

Changes:

• Inline HndWriteBarrierWorker into src/coreclr/gc/handletable.inl and remove the out-of-line implementation from handletable.cpp.
• Add an early-return fast-path in the worker to skip HandleFetchType(handle) when GetConvertedGeneration(value) is 0.
• Move/adjust declarations (GetConvertedGeneration, HandleFetchType) to be available to the inlined implementation and update includes accordingly.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
src/coreclr/gc/handletablepriv.h	Removes now-unneeded forward declarations previously used by the out-of-line worker.
src/coreclr/gc/handletable.inl	Adds the inlined HndWriteBarrierWorker implementation and updates HndAssignHandle to call it directly.
src/coreclr/gc/handletable.h	Adds handletableconstants.h include and exposes needed non-DAC prototypes for the inlined worker.
src/coreclr/gc/handletable.cpp	Removes the out-of-line HndWriteBarrierWorker implementation (now provided inline).

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated no new comments.

Suppressed comments (1)

src/coreclr/gc/handletable.inl:68

• The comment describing async pinned handle behavior has a few grammar/wording issues (e.g., missing space after //, “need”/“consider”). Updating it will make the intent clearer without changing behavior.

        //OverlappedData need special treatment: because all user data pointed by it needs to be reported by this handle,
        //its age is consider to be min age of the user data, to be simple, we just make it 0

[Copilot]

Pull request overview

Copilot reviewed 10 out of 10 changed files in this pull request and generated 1 comment.

Suppressed comments (2)

src/coreclr/gc/handletablepriv.h:171

• HandleFetchType in handletable.inl reads the block type via raw uint8_t* indexing (segment[HANDLE_SEGMENT_BLOCK_TYPE_OFFSET + uBlock]). The existing offsetof static_assert guards the offset, but the implementation also assumes rgBlockType elements are 1 byte. Adding a size check here makes that assumption explicit and prevents silent breakage if the field type ever changes.

static_assert(offsetof(_TableSegmentHeader, rgBlockType) == HANDLE_SEGMENT_BLOCK_TYPE_OFFSET);

src/coreclr/gc/handletable.inl:43

• New code in handletable.inl uses C-style casts when deriving segment pointers (e.g., (uint8_t*)((uintptr_t)handle & HANDLE_SEGMENT_ALIGN_MASK)), while nearby code already uses reinterpret_cast. This is easy to fix and aligns with the repo's native guidance to prefer C++ casts.

    uint8_t* barrier = (uint8_t*)((uintptr_t)handle & HANDLE_SEGMENT_ALIGN_MASK);
    _ASSERTE(barrier);

    uintptr_t offset = (uintptr_t)handle & HANDLE_SEGMENT_CONTENT_MASK;
    _ASSERTE(offset >= HANDLE_HEADER_SIZE);

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated no new comments.

Suppressed comments (1)

src/coreclr/gc/handletable.inl:70

• Fix grammar/spacing in this comment (missing space after //, subject/verb agreement, and "consider" → "considered") to keep comments clear and consistent.

        //OverlappedData need special treatment: because all user data pointed by it needs to be reported by this handle,
        //its age is consider to be min age of the user data, to be simple, we just make it 0

[jkotas]

@EgorBot -windows_x64 -linux_x64 -linux_arm64

using System.Runtime.InteropServices;
using BenchmarkDotNet.Attributes;

public class Bench
{
    [Benchmark]
    public void GCHandle()
    {
        using (var hnd = new PinnedGCHandle<string>("Hello"))
        {
        }
    }
}

[jkotas]

@copilot Delete TableFreeHandlesToCache method - it is dead code

[jkotas]

@copilot Make sure that TableFreeSingleHandleToCache is inlined in the one place it is called

[jkotas]

@copilot GetGlobalHandleStore never fails. Delete unnecessary error handling in appdomain.cpp

[jkotas]

@copilot Make sure that GetGlobalHandleStore iis cached in a static in NativeAOT - similar to how it is done in CoreCLR

[Copilot]

@copilot Delete TableFreeHandlesToCache method - it is dead code

Deleted the dead implementation and declaration in 796149b. The CoreCLR runtime build succeeds with 0 warnings and 0 errors.

Note

This comment was generated by AI/Copilot.

[Copilot]

Pull request overview

Copilot reviewed 10 out of 10 changed files in this pull request and generated no new comments.

Suppressed comments (1)

src/coreclr/gc/handletable.inl:71

• The comment for async pinned handles has grammar issues and is a bit hard to read (e.g., "need"/"is consider"), which makes the rationale unclear in a subtle write-barrier path. Consider rewording for clarity while keeping the intent the same.

        //OverlappedData need special treatment: because all user data pointed by it needs to be reported by this handle,
        //its age is consider to be min age of the user data, to be simple, we just make it 0
        if (uType == HNDTYPE_ASYNCPINNED)

[Copilot]

@copilot Make sure that GetGlobalHandleStore iis cached in a static in NativeAOT - similar to how it is done in CoreCLR

Added g_pGlobalHandleStore in 55c837b, set alongside g_pGCHandleManager in both GCHeapUtilities::InitializeDefaultGC and InitializeStandaloneGC, and exposed via GCHandleUtilities::GetGlobalHandleStore(). Updated HandleTableHelpers.cpp to use it. Verified clr.nativeaotruntime builds with 0 errors.

@copilot GetGlobalHandleStore never fails. Delete unnecessary error handling in appdomain.cpp

Removed the null check and COMPlusThrowOM in 55c837b.

@copilot Make sure that TableFreeSingleHandleToCache is inlined in the one place it is called

Addressed in 55c837b — moved TableFreeSingleHandleToCache into handletablepriv.h as an __inline function (guarded by #ifndef DACCESS_COMPILE since it uses Interlocked::ExchangePointer in a way that doesn't compile for DAC's OBJECTHANDLE type), so it's inlined at its only call site in HndDestroyHandle. Verified clr.runtime builds with 0 errors, including cee_dac.

[Copilot]

Pull request overview

Copilot reviewed 15 out of 15 changed files in this pull request and generated 1 comment.