Use Reflection.Emit accessors when dynamic code is supported

[eiriktsarpalis]

This restores the Reflection.Emit-based member accessor whenever dynamic code is supported, reverting only the accessor-selection change from #130503. The reflection invocation fixes and exception-propagation tests from that PR remain intact.

Performance evidence

The accessor-selection change in #130503 was the first bad commit for dotnet/perf-autofiling-issues#76620. On Ubuntu 22.04 x64 with WASM and JavaScriptCore:

Commit	Reflection	Source generation	Allocation
Parent 8df98c4	6.788 us	3.804 us	48 B
#130503 (ddb840c)	14.115 us	3.777 us	128 B

The reflection-based serializer became 2.08x slower while the source-generation control remained unchanged. Additional JavaScriptCore and V8 benchmarks found interpreted DynamicMethod accessors 3–41x faster per repeated invocation than reflection. Creating them was 62–66x slower, but System.Text.Json caches generated accessors and amortizes that cost.

Tradeoff

This intentionally reverses the trimming decision made for #38693. Interpreter deployments will retain the Reflection.Emit stack, restoring the approximately 50 KB size cost measured there and increasing accessor creation cost. The measurements above show that interpreted emitted IL does provide a substantial steady-state throughput and allocation benefit for reflection-based serialization, contrary to the premise of #130503.

Validation

• dotnet build
• dotnet build /t:test .\tests\System.Text.Json.Tests\System.Text.Json.Tests.csproj — 53,409 .NET tests and 53,163 .NET Framework tests passed with zero failures

Note

This pull request was generated by GitHub Copilot.

[azure-pipelines]

Azure Pipelines:
Successfully started running 3 pipeline(s).
13 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[dotnet-policy-service]

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

[Copilot]

Pull request overview

This PR changes System.Text.Json’s internal MemberAccessor selection logic to prefer the Reflection.Emit-based accessor whenever dynamic code is supported (rather than requiring dynamic code to be compiled), restoring the pre-#130503 behavior to improve steady-state reflection-based serialization performance on interpreter-based runtimes.

Changes:

• Switch accessor selection from RuntimeFeature.IsDynamicCodeCompiled back to RuntimeFeature.IsDynamicCodeSupported.
• Update/remove the prior rationale comment (replacing it with a shorter comment).

[jkotas]

contrary to the premise of #130503

Every other perf improvement without numbers turns out to be a regression

[MichalStrehovsky]

contrary to the premise of #130503

Every other perf improvement without numbers turns out to be a regression

This one actually had perf numbers saying this is a throughput regression: #38693 (comment) and #38693 (comment). So it's more like "AI can help us to churn out fixes without reading the past conversations that were saying doing this without a tradeoff is hard".

[eiriktsarpalis]

Well, my takeaway is that automation (AI and performance testing in this case) at the end of the day steered us towards the right direction with minimal human involvement. And we fixed a few lurking bugs along the way too.