Use reader/writer lock instead of mutex for ComWrappers RCW cache

[AaronRobinsonMSFT]

Switching to a reader/writer lock improves throughput for faster RCW look up by 4x. This was observed in a private repro associated with #90964.

Fixes #90964

[AaronRobinsonMSFT]

/cc @jkotas @dotnet/interop-contrib @manodasanW

[jkotas]

Would it be possible to extract a repro into a microbenchmark in https://github.com/dotnet/performance?

Also, it would be good to check how native AOT does on this microbenchmark.

[AaronRobinsonMSFT]

Would it be possible to extract a repro into a microbenchmark in https://github.com/dotnet/performance?

I will need to construct one. The repro was using C#/WinRT and would be non-trivial to recreate. I will see what I can do.

Also, it would be good to check how native AOT does on this microbenchmark.

The data structure being used in AOT is the ConditionalWeakTable. I don't think this has the same mutex issues that we had in CoreCLR.

[manodasanW]

Also, it would be good to check how native AOT does on this microbenchmark.

The data structure being used in AOT is the ConditionalWeakTable. I don't think this has the same mutex issues that we had in CoreCLR.

For AOT, thoughts on the lock here for rcwCache.

[AaronRobinsonMSFT]

Also, it would be good to check how native AOT does on this microbenchmark.

The data structure being used in AOT is the ConditionalWeakTable. I don't think this has the same mutex issues that we had in CoreCLR.

For AOT, thoughts on the lock here for rcwCache.

I was misinformed. Sad. Okay, that needs to change.

[AaronRobinsonMSFT]

Would it be possible to extract a repro into a microbenchmark in https://github.com/dotnet/performance?

Also, it would be good to check how native AOT does on this microbenchmark.

@jkotas and @manodasanW I'm genuinely shocked at this, but converting NAOT to use a reader/writer lock makes NAOT slower for a microbenchmark. I was able to see the expected speed in the microbenchmark with CoreCLR and the reader/writer lock so it is capturing the scenario, but there is nothing to do here for NAOT.

I will work on pushing the benchmark into https://github.com/dotnet/performance.

[AaronRobinsonMSFT]

NAOT

native_baseline\Perf.exe
Parallel.ForEach: 409 ms.
native\Perf.exe
Parallel.ForEach: 520 ms.

CoreCLR

Perf_baseline.cmd
Parallel.ForEach: 2060 ms.
Perf.cmd
Parallel.ForEach: 428 ms.

[AaronRobinsonMSFT]

Benchmark PR - dotnet/performance#3298

[AaronRobinsonMSFT]

/backport to release/8.0

[github-actions]

Started backporting to release/8.0: https://github.com/dotnet/runtime/actions/runs/5981298027