[tests] Add Android interop benchmarks and reduce JNI overhead

[simonrozsival]

Summary

• add focused BenchmarkDotNet coverage for trimmable typemap lookups, peer activation/cache behavior, JNI-to-Java-to-UCO roundtrips, Android APIs, object returns, and primitive/jagged arrays
• decompose Uri.Parse() into Java execution, string creation, and managed peer activation
• eliminate transient JNI array-signature allocations by using UTF-8 literals for built-in array classes
• avoid redundant class-name discovery for known jagged-array element types and use IsInstanceOf() for Java-to-managed validation
• resolve sealed framework peer proxies from the requested managed type instead of materializing the runtime Java class name
• cache raw-JNI benchmark method metadata outside the measured operation

Samsung S23 results

Benchmark	Before	After	Allocation before	Allocation after
byte[256] conversion	956.9 ns	515.2 ns	336 B	280 B
four-element byte[256][] conversion	6.010 µs	2.940 µs	1,432 B	1,176 B

The remaining allocations are exactly the required managed array representations: 280 B for byte[256], and 1,176 B for the outer array plus four inner arrays. The raw JNI URI controls now allocate 0 B.

For sealed framework peer activation, the existing runtime class discovery components cost 471.8 ns and 112 B (GetObjectClass plus GetJniTypeName). Validating the already-known target class directly costs 384.9 ns and 0 B, removing the class-name allocation and reducing this lookup stage by about 18%.

Validation

• built and ran the targeted benchmark filters on a Samsung S23 (SM-S911B)
• built the complete Android benchmark project in Release with CoreCLR and the trimmable typemap
• built Mono.Android with the sealed-peer optimization
• added compatible, incompatible, and sealed closed-generic peer activation coverage
• existing enum, nullable primitive, byte, sbyte, and jagged-array coverage exercises the changed conversion paths

[Copilot]

Copilot review overview

🔵 Needs a closer look

It modifies core JNIEnv interop array-class/type-compat logic in addition to adding benchmarks, making the blast radius broad enough to warrant final human review.

Review tier: Lite
Findings: 1

New issues introduced by this change (1)

Severity	Finding
	tests/​Android.Benchmarks/​AndroidArrayMarshallingBenchmarks.cs — 💡 suggestion Operational implications — This benchmark uses Android 12 / API 31 types (e.g.,…

What changed in this PR

This PR extends the tests/Android.Benchmarks project with targeted BenchmarkDotNet scenarios for common interop and framework usage patterns, and makes a small runtime-side optimization in JNIEnv to reduce transient allocations during array-class lookups (notably for built-in primitive/boxed array signatures).

Changes:

• Add new BenchmarkDotNet benchmark types covering typemap lookups, peer activation/cache behavior, raw JNI roundtrips, object creation, and array marshalling.
• Decompose URI parsing into separate “bound” vs “raw JNI” vs “activation from prepared JNI handles” benchmarks to isolate allocation sources.
• Optimize JNIEnv array type compatibility checks and class lookup paths by using UTF-8 literals for built-in array class names and a fast-path for jagged array converter selection.

File	Description
tests/​Android.Benchmarks/​UriParseAllocationBenchmarks.cs	Adds URI parse benchmarks isolating string creation, raw JNI call, and managed peer activation allocations.
tests/​Android.Benchmarks/​TypeMapLookupBenchmarks.cs	Adds benchmarks for typemap type ↔ signature resolution (framework/app/missing).
tests/​Android.Benchmarks/​PeerLookupBenchmarks.cs	Adds benchmarks for GetObject<T> peer cache hit/miss and global-ref scenarios.
tests/​Android.Benchmarks/​PeerActivationComponentBenchmarks.cs	Adds microbenchmarks for common activation components (class lookup/name/hash/global ref).
tests/​Android.Benchmarks/​ExportRoundtripBenchmarks.cs	Adds JNI→managed exported method roundtrip benchmarks (primitive + string).
tests/​Android.Benchmarks/​AndroidObjectCreationBenchmarks.cs	Adds benchmarks for common framework object creation APIs (Intent, Bitmap, etc.).
tests/​Android.Benchmarks/​AndroidArrayMarshallingBenchmarks.cs	Adds API 31+ array marshalling benchmarks (AppSearch, Parcel, Base64, digest, prepared JNI arrays).
tests/​Android.Benchmarks/​AndroidApiBenchmarks.cs	Adds benchmarks for frequently-used Android API calls (Bundle, resources, system services, view ops).
tests/​Android.Benchmarks/​Android.Benchmarks.csproj	Forces CoreCLR (UseMonoRuntime=false) and trimmable typemap for the benchmarks app.
src/​Mono.Android/​Android.Runtime/​JNIEnv.cs	Reduces array-signature allocations via UTF-8 built-in array class names and adjusts array converter selection/compat checks.

[jonathanpeppers]

/review

[github-actions]

✅ Android PR Reviewer completed successfully!

Generated by Android PR Reviewer for #12644

[github-actions]

Reviewed the runtime array-class/compatibility changes, the sealed framework-peer fast path, and the added tests and benchmarks. I found no additional actionable code issues beyond the existing API 31 benchmark-execution comment.

CI build #1579150 is still in progress and currently red: Package Tests macOS > Tests > APKs 1 has failed, with eight lanes still running. The Azure CLI could not initialize its configuration in this environment, so I could not verify the failure's root cause and am not attributing it to this change.

Generated by Android PR Reviewer for #12644 · gpt56 · 199.8 AIC · ⌖ 8.83 AIC · ⊞ 25.7K
_{Comment /review to run again}