Lack of devirtualisation on Bindable<T>.Value

[smoogipoo]

We are losing a considerable amount of performance due to Bindable<T>.Value being virtual. Making BindableInt override + seal works however I'd rather push towards improving the bindable structure further instead since this sounds quite limiting.

Current:

Method	Mean	Error	StdDev	Allocated
BenchmarkGetValue	16.25 us	0.010 us	0.009 us	-

With a sealed override:

Method	Mean	Error	StdDev	Allocated
BenchmarkGetValue	2.871 us	0.0231 us	0.0193 us	-

With referencing the field directly:

Method	Mean	Error	StdDev	Allocated
BenchmarkGetValue	2.837 us	0.0548 us	0.0539 us	-

[peppy]

I think there's quite a few scenarios we need to use sealed, too.

[smoogipoo]

The bindable interfaces are also troublesome, since every interface method access is a virtual call regardless of whether we devirtualise Value.

C# doesn't yet have interface devirtualisation:

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
BenchmarkNormalClass	2.859 us	0.0013 us	0.0012 us	-	-	-	-
BenchmarkNormalClassAsInterface	21.643 us	0.0011 us	0.0010 us	-	-	-	-
BenchmarkClassWithVirtual	13.534 us	0.0014 us	0.0013 us	-	-	-	-
BenchmarkClassWithVirtualAsInterface	21.644 us	0.0006 us	0.0005 us	-	-	-	-
BenchmarkSealedClass	2.855 us	0.0110 us	0.0097 us	-	-	-	-
BenchmarkSealedClassAsInterface	21.661 us	0.0276 us	0.0230 us	-	-	-	-
BenchmarkStruct	2.711 us	0.0005 us	0.0005 us	-	-	-	-
BenchmarkStructAsInterface	35.178 us	0.0207 us	0.0173 us	-	-	-	-

Perhaps a ReadOnlyBindable class is something to consider.

[turbedi]

I tried your benchmark on my machine with .NET 6, but also Dynamic PGO enabled:

Method	Job	Mean	Error	StdDev	Median
BenchmarkNormalClass	Default mode	2.400 μs	0.0139 μs	0.0123 μs	2.402 μs
BenchmarkNormalClass	Dynamic PGO	2.422 μs	0.0102 μs	0.0096 μs	2.422 μs
BenchmarkNormalClassAsInterface	Default mode	18.929 μs	0.0435 μs	0.0386 μs	18.923 μs
BenchmarkNormalClassAsInterface	Dynamic PGO	5.174 μs	0.0041 μs	0.0037 μs	5.174 μs
BenchmarkClassWithVirtual	Default mode	14.211 μs	0.0287 μs	0.0255 μs	14.207 μs
BenchmarkClassWithVirtual	Dynamic PGO	5.089 μs	0.1001 μs	0.1644 μs	5.167 μs
BenchmarkClassWithVirtualAsInterface	Default mode	18.884 μs	0.0326 μs	0.0272 μs	18.880 μs
BenchmarkClassWithVirtualAsInterface	Dynamic PGO	5.110 μs	0.0997 μs	0.1430 μs	5.166 μs
BenchmarkSealedClass	Default mode	2.399 μs	0.0140 μs	0.0131 μs	2.396 μs
BenchmarkSealedClass	Dynamic PGO	2.419 μs	0.0055 μs	0.0046 μs	2.421 μs
BenchmarkSealedClassAsBase	Default mode	14.271 μs	0.0366 μs	0.0342 μs	14.291 μs
BenchmarkSealedClassAsBase	Dynamic PGO	5.125 μs	0.1056 μs	0.1335 μs	5.178 μs
BenchmarkSealedClassAsInterface	Default mode	18.998 μs	0.0197 μs	0.0164 μs	19.004 μs
BenchmarkSealedClassAsInterface	Dynamic PGO	4.769 μs	0.0048 μs	0.0037 μs	4.769 μs
BenchmarkStruct	Default mode	2.417 μs	0.0165 μs	0.0154 μs	2.420 μs
BenchmarkStruct	Dynamic PGO	2.413 μs	0.0093 μs	0.0082 μs	2.416 μs
BenchmarkStructAsInterface	Default mode	32.995 μs	0.0312 μs	0.0291 μs	32.990 μs
BenchmarkStructAsInterface	Dynamic PGO	4.775 μs	0.0244 μs	0.0216 μs	4.770 μs

According to the post Dynamic PGO can also cause regressions in some cases, but this is supposed to be fixed in .NET 7

[smoogipoo]

Good to know. I believe in general we can turn on Dynamic PGO with .NET 6 already, did some rough benchmarks as such a while back. Will need to see if there are runtimeconfig flags for it, or if we'd need to wait for it to be turned on by default.

[turbedi]

Runtime configs are planned for .NET 7

[Rekkonnect]

Since the project was upgraded to .NET 8.0, maybe revisiting this is a good idea

[smoogipoo]

Dynamic PGO is enabled by default in .NET 8.