Avoid an allocation on every GetOrCreateComInterfaceForObject call

[Sergio0694]

ComWrappers.GetOrCreateComInterfaceForObject passes the state its ConditionalWeakTable.GetOrAdd factory needs as an anonymous type:

ManagedObjectWrapperHolder managedObjectWrapper = _managedObjectWrapperTable.GetOrAdd(instance, static (c, state) =>
{
    ManagedObjectWrapper* value = state.This.CreateManagedObjectWrapper(c, state.flags);
    return new ManagedObjectWrapperHolder(value, c);
}, new { This = this, flags });

Anonymous types are classes, so that's 32 bytes allocated on every call. The factory only runs when the object doesn't have a wrapper yet, and in practice almost every call finds one that already exists, so nearly all of it is garbage that never gets used. This is the path every managed object takes on its way into native code, so it adds up.

GetOrAdd<TArg> doesn't constrain TArg (it even allows ref structs), so a small struct works here and costs nothing to pass:

private readonly struct CreateManagedObjectWrapperState(ComWrappers comWrappers, CreateComInterfaceFlags flags)
{
    public ComWrappers ComWrappers { get; } = comWrappers;

    public CreateComInterfaceFlags Flags { get; } = flags;
}

No behaviour change, no API change.

Benchmarks

Release runtime, x64, two alternating passes per build. The allocation numbers were identical across runs.

Scenario	main	this PR
Wrapper already exists, 8000 distinct objects	90.30 ns, 32 B	71.18 ns, 0 B	−21%
Wrapper already exists, same object each time	37.35 ns, 32 B	34.37 ns, 0 B	−8%
Wrapper has to be created	598.4 ns, 138 B	575.7 ns, 106 B	−4%

The lookup path is allocation free now. The first row gains more than the second because the garbage was also costing GC pressure and cache misses, not just the allocation itself. The creation row is within noise on time, since it's dominated by building the wrapper, but it still drops the 32 bytes.

Testing

ComWrappersTests, ComWrappersTestsBuiltInComDisabled, GcRestrictedCalloutReversePInvoke, GlobalInstanceMarshallingTests, GlobalInstanceMarshallingTestsBuiltInComDisabled, GlobalInstanceTrackerSupportTests_TargetWindows and WeakReferenceTest all pass against the modified CoreLib, built Checked.

Note

Parts of this pull request description were generated with GitHub Copilot. All benchmark numbers in it were measured locally.

[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: @dotnet/interop-contrib
See info in area-owners.md if you want to be subscribed.

[jkoritzinsky]

It's really disappointing that we can't use anonymous types here TBH. Effectively makes things overly wordy when it really shouldn't be necessary.

LGTM

[Copilot]

Pull request overview

This PR removes a per-call heap allocation in ComWrappers.GetOrCreateComInterfaceForObject by replacing an anonymous-type state object passed to ConditionalWeakTable.GetOrAdd with a small readonly struct state payload.

Changes:

• Replace new { This = this, flags } with CreateManagedObjectWrapperState to avoid allocating state on every call.
• Update the GetOrAdd value-factory to use the new state shape (state.ComWrappers, state.Flags).
• Add a private nested state struct (with documentation) to carry ComWrappers and CreateComInterfaceFlags.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

[MattParkerDev]

It's really disappointing that we can't use anonymous types here TBH. Effectively makes things overly wordy when it really shouldn't be necessary.

It could be a tuple instead?

[Sergio0694]

@MattParkerDev we tried using value tuples before in ComWrappers and we reverted it. It roots a bunch of additional cruft that can't be trimmed and was causing something like a 7KB size regression on NAOT for no reason. A custom struct is better.