Optimize JIT_ClassProfile32

[EgorBo]

I'm still seeing JIT_ClassProfile32 being top5 hottest functions in the TE traces + was surprised to see it for Paint.NET's traces (forced to stay in tier0-instrumented).

New codegen:

This PR does:

1. optimizes classProfile->Count++ operation. volatile seems unnecessary here
2. force inline CheckSample into JIT_ClassProfile32
3. Little massage to produce better codegen on MSVC (other compilers do a better job as is)
4. Replaces pMT->GetLoaderAllocator()->IsCollectible() with pMT->Collectible()

(mostly written by @jakobbotsch)

Open questions

1. It seems like we need to collect Native PGO with DOTNET_TieredPGO=1 to produce a better NPGO coverage.

[jakobbotsch]

LGTM! BTW, when I looked at it the FORCEINLINE was not required for MSVC to inline CheckSample, but still seems fine to me to have it given that we know it's very hot.

Would also like to know what the intention of the volatile increment was since that is being removed (cc @AndyAyersMS)

[AndyAyersMS]

We might also look at implementing something closer to the true reservoir sampling via if ((x % count) >= S) or if you don't want that unknown value modulus, at least decreasing the probability of an update once count is large.

Would also like to know what the intention of the volatile increment was since that is being removed (cc @AndyAyersMS)

Doesn't look like it is needed. I don't remember why I had it that way.

[AndyAyersMS]

@EgorBo did you collect any data for this one?

[EgorBo]

@EgorBo did you collect any data for this one?

I did, the improvement was quite small, I assume, mostly because the main issue is in classProfile->Count++ cache contention

[EgorBo]

/azp list

[EgorBo]

/azp run runtime-coreclr pgo, runtime-coreclr libraries-pgo

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).

[EgorBo]

Failure is #82275