Enable RuntimeType.GenericCache to hold multiple types of cache entries

[jkoritzinsky]

Enable GenericCache to hold multiple types of cache entries in preparation for some boxing optimizations that can use the cache. Try to encapsulate the design so the users of the GenericCache don't need to think about multiple cache entries.

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-system-runtime
See info in area-owners.md if you want to be subscribed.

[jkoritzinsky]

/benchmark micro aspnet-perf-win runtime --variable filter=System.Tests.Perf_Array

[jkoritzinsky]

/benchmark micro aspnet-perf-win runtime --variable filter=System.Tests.Perf_Enum

[jkoritzinsky]

/benchmark micro aspnet-perf-win runtime --variable filter=System.Reflection.Activator

[pr-benchmarks]

Benchmark started for micro on aspnet-perf-win with runtime and arguments --variable filter=*System.Reflection.Activator*. Logs: link

[pr-benchmarks]

micro - aspnet-perf-win

| benchmark                                    | mean (micro.base) | mean (micro.pr) | ratio | allocated (micro.base) | allocated (micro.pr) | ratio |
| -------------------------------------------- | ----------------- | --------------- | ----- | ---------------------- | -------------------- | ----- |
| Activator<EmptyClass>.CreateInstanceGeneric  |          7.302 ns |       70.649 ns |  9.67 |                   24 B |                 56 B |  2.33 |
| Activator<EmptyClass>.CreateInstanceNames    |            1.5 μs |          1.5 μs |  1.00 |                  920 B |                920 B |  1.00 |
| Activator<EmptyClass>.CreateInstanceType     |          6.434 ns |        8.138 ns |  1.26 |                   24 B |                 24 B |  1.00 |
| Activator<EmptyStruct>.CreateInstanceGeneric |         0.0001 ns |       0.0000 ns |  0.20 |                    0 B |                  0 B |       |
| Activator<EmptyStruct>.CreateInstanceNames   |            1.3 μs |          1.3 μs |  1.04 |                  744 B |                744 B |  1.00 |
| Activator<EmptyStruct>.CreateInstanceType    |          5.630 ns |        6.893 ns |  1.22 |                   24 B |                 24 B |  1.00 |

[pr-benchmarks]

Benchmark started for micro on aspnet-perf-win with runtime and arguments --variable filter=*System.Tests.Perf_Enum*. Logs: link

[jkotas]

I have run this micro-benchmark on Win x64 to see the perf of composite cache:

using System;
using System.Diagnostics;
using System.Runtime.CompilerServices;

var sw = new Stopwatch();
for (;;)
{
   sw.Restart();
   Work();
   Internal.Console.WriteLine(sw.ElapsedMilliseconds.ToString());
}

static void Work()
{
    for (int i = 0; i < 100000000; i++)
    {
        GC.KeepAlive(Activator.CreateInstance(typeof(object)));
        GC.KeepAlive(RuntimeHelpers.GetUninitializedObject(typeof(object)));
    }
}

It runs about 1.3x-1.4x slower compared to current main. These seems to be a lot of perf lost in the generics. Can we do better?

[jkoritzinsky]

I'll see if I can make it faster.

I've been focusing on matching performance with our existing (single-entry-focused) microbenchmarks. I've been able to get within single-digit percentage (even with a possible improvement in one run) for Activator.CreateInstance. I'll spend some time on the composite case.

[jkotas]

I have tried to replace the InlineArray with explicit fields jkotas@1a96a78 . It fixed the perf for my micro-benchmark. (The change is not complete - it is missing handling of EnumInfo.)