JsonSerializerOptions.MemberAccessorStrategy shouldn't use Reflection.Emit when IsDynamicCodeCompiled is false

[eerhardt]

When trimming a Blazor WASM app, the last usage of Reflection.Emit (after fixing #38678) is coming from System.Text.Json.Serialization.JsonSerializerOptions:

runtime/src/libraries/System.Text.Json/src/System/Text/Json/Serialization/JsonSerializerOptions.cs

Lines 423 to 438 in d89772a

	internal MemberAccessor MemberAccessorStrategy
	{
	get
	{
	if (_memberAccessorStrategy == null)
	{
	#if NETFRAMEWORK || NETCOREAPP
	_memberAccessorStrategy = new ReflectionEmitMemberAccessor();
	#else
	_memberAccessorStrategy = new ReflectionMemberAccessor();
	#endif
	}
	return _memberAccessorStrategy;
	}
	}

However, on Mono WASM, RuntimeFeature.IsDynamicCodeCompiled is always false, so using Reflection.Emit is probably a waste, and it brings in a decent amount of code. In my investigations I find it removing ~50KB of IL if we trim this usage of Reflection.Emit.

We should change this code to something more like:

 internal MemberAccessor MemberAccessorStrategy
 {
 get
 {
 if (_memberAccessorStrategy == null)
 {
#if NETFRAMEWORK || NETCOREAPP
 if (RuntimeFeature.IsDynamicCodeCompiled)
 {
 _memberAccessorStrategy = new ReflectionEmitMemberAccessor();
 }
 else
 {
 _memberAccessorStrategy = new ReflectionMemberAccessor();
 }
#else
 _memberAccessorStrategy = new ReflectionMemberAccessor();
#endif
 }

 return _memberAccessorStrategy;
 }
 }

With changing the code to the above, on a default template Blazor WASM app, I am seeing size savings of:

Build	Size
master	3,366,912 bytes
#38729	3,039,232 bytes
#38729 + this change	2,990,080 bytes

So almost a 50 KB savings by allowing the removing all usages of System.Reflection.Emit.

cc @steveharter @layomia @vitek-karas @marek-safar

[steveharter]

FWIW we previously used RuntimeFeature.IsDynamicCodeSupported but that API requires netstandard 2.1, and we only had support for 2.0 in STJ.dll and due to the switch to netcoreapp (instead of netstandard) it was decided not worth the hit of a larger package just for this.

But using an ifdef branch for netcoreapp that checks RuntimeFeature should work great.

[eerhardt]

I ran the attached JSON perf application with and without this change.

Program (1).zip

5.0.0-preview.8.20354.5

serialize	deserialize	writer
Serialization took 1509 ms	Deserialization took 4619 ms	Writer took 1252 ms
Serialization took 1548 ms	Deserialization took 4690 ms	Writer took 1272 ms
Serialization took 1537 ms	Deserialization took 4743 ms	Writer took 1224 ms

Proposed change

serialize	deserialize	writer
Serialization took 1843 ms	Deserialization took 5391 ms	Writer took 1196 ms
Serialization took 1832 ms	Deserialization took 5364 ms	Writer took 1196 ms
Serialization took 1847 ms	Deserialization took 5338 ms	Writer took 1185 ms

So it is ~15-20% slower to use the Reflection based strategy after everything is warmed up. (Note this is explicitly not testing the time it takes to emit IL during warmup.)

[eerhardt]

Since the performance tradeoff of taking this change isn't necessarily a net-win, we won't be making this change in 5.0. Moving to 6.0.