[Mono][Perf] JIT Performance Regression, multiple benchmarks, 7 % to 909 %

[naricc]

This is a master issue to track a performance regression effecting several benchmarks that seem related.

The date range suggests that the cause is one of the commits in this range: 24d5043...7664a9a of which, only c723f5e is plausible.

These are the effected benchmarks on their auto-filed issues:



JIT
[Perf -20%] PerfLabTests.LowLevelPerf.GenericClassGenericStaticField
https://github.com/DrewScoggins/performance-2/issues/2887


[Perf -11%] PerfLabTests.LowLevelPerf.NewDelegateClassEmptyStaticFn 
https://github.com/DrewScoggins/performance-2/issues/2996

[Perf -18%] System.Threading.Tests.Perf_Lock.ReaderWriterLockSlimPerf 
DrewScoggins/performance-2#2709

[Perf -27%] System.Xml.Linq.Perf_XDocument.Create 
DrewScoggins/performance-2#2712

[Perf -9%] System.Tests.Perf_UInt16.ToString
DrewScoggins/performance-2#2707

[Perf -13%] Burgers.Test0 
DrewScoggins/performance-2#2995

[Perf -14%] System.Threading.Tasks.ValueTaskPerfTest.CreateAndAwait_FromYieldingAsyncMethod
DrewScoggins/performance-2#2703

[Perf -10%] System.Numerics.Tests.Perf_VectorConvert.Convert_uint_float
DrewScoggins/performance-2#2994

[Perf -12%] System.Threading.Channels.Tests.SpscUnboundedChannelPerfTests.PingPong
DrewScoggins/performance-2#2993

[Perf -13%] Devirtualization.Boxing.InterfaceTypeCheckAndCall
DrewScoggins/performance-2#2724

[Perf -12%] System.Threading.Tasks.Tests.Perf_AsyncMethods.SingleYieldMethodInvocation
DrewScoggins/performance-2#2992

[Perf -9%] System.Tests.Perf_Array.ArrayAssign3D 
DrewScoggins/performance-2#2991

DrewScoggins/performance-2#2715
[Perf -32%] System.Tests.Perf_Object.ctor 

[Perf -7%] Benchstone.BenchF.BenchMk2.Test 
DrewScoggins/performance-2#2990

[Perf -13%] System.Numerics.Tests.Perf_Vector3.SquareRootBenchmark
DrewScoggins/performance-2#2989

DrewScoggins/performance-2#2720
[Perf -15%] System.Threading.Channels.Tests.BoundedChannelPerfTests.PingPong

DrewScoggins/performance-2#2706
[Perf -193%] System.Threading.Tests.Perf_Timer.ShortScheduleAndDisposeWithFiringTimers

DrewScoggins/performance-2#2717
[Perf -9%] Benchstone.BenchF.Simpsn.Test

[Perf -6%] System.Memory.Constructors.ReadOnlySpanImplicitCastFromArraySeg
https://github.com/DrewScoggins/performance-2/issues/2988

[Perf -809%]

System.Threading.Tests.Perf_ThreadStatic.GetThreadStatic
https://github.com/DrewScoggins/performance-2/issues/2711



[Perf -9%] System.Net.Primitives.Tests.IPAddressPerformanceTests.Ctor_Bytes
DrewScoggins/performance-2#2987



[Perf -15%] Microsoft.Extensions.DependencyInjection.GetServiceIEnumerable.Transient
DrewScoggins/performance-2#2716


[Perf -17%] System.Numerics.Tests.Perf_Vector4.SquareRootBenchmark
https://github.com/DrewScoggins/performance-2/issues/2696



[Perf -7%] System.Collections.Tests.Perf_BitArray (2) 
https://github.com/DrewScoggins/performance-2/issues/2700

Interprter:


https://github.com/DrewScoggins/performance-2/issues/2887


[Perf -20%] PerfLabTests.LowLevelPerf.GenericClassGenericStaticField

[Dotnet-GitSync-Bot]

I couldn't figure out the best area label to add to this issue. If you have write-permissions please help me learn by adding exactly one area label.

[ghost]

Tagging subscribers to this area:
See info in area-owners.md if you want to be subscribed.

[naricc]

@vargaz It looks like the change to building with cmake may be responsible for this regression, based on the dates.

[vargaz]

What platform is this ?

[naricc]

@vargaz ubuntu x64

[vargaz]

Can't reproduce this locally, a lot of these benchmarks are very simple, the first one is:
for (int i = 0; i < n_iter; i++)
GenericClass.aGenericStaticFieldT = 1;
which just gets compiled into a simple loop.

[vargaz]

Can somebody reproduce this locally ?

[naricc]

@vargaz What is the issue with reproducing it? You were not able to run the benchmarks, or you did and it didn't repro? System.Threading.Tests.Perf_Timer.ShortScheduleAndDisposeWithFiringTimers seems like the biggest difference, so it should be the easiest to see, if it is a real issue.

[naricc]

I am not sure what cause this, but it seems to have resolved itself.