[TrimmableTypeMap][NativeAOT] Auto-detect and scan runtime-only assemblies (fixes stripped runtime ACWs)

[simonrozsival]

Summary

Under the trimmable typemap on NativeAOT, the runtime host assembly Microsoft.Android.Runtime.NativeAOT — and its only Java Callable Wrapper type, UncaughtExceptionMarshaler — was never scanned by the typemap generator. Its JCW, typemap entry, and acw-map entry were all missing, so R8 had nothing to keep, the ACW was dropped from classes.dex, and the app crashed at startup in JavaInteropRuntime.init (setDefaultUncaughtExceptionHandler).

More generally: any runtime-only assembly with Java peers was invisible to the typemap generator. This fixes the whole class of problem, not just UncaughtExceptionMarshaler.

Root cause

_GenerateTrimmableTypeMap runs in the RID-independent outer build over @(ReferencePath) (the compile closure: app + references + ref-pack framework assemblies). Runtime-only assemblies have no ref-pack counterpart and are only pulled from the RID-specific runtime pack in the per-RID inner build, so they never enter the generator's input set. Runtime-pack resolution (ResolveRuntimePackAssets) fundamentally requires a single RuntimeIdentifier.

Fix — auto-detect via the SDK's own resolution

No hard-coded assembly names, no per-assembly reference-assembly plumbing:

• _ResolveRuntimeOnlyAssembliesForTypeMap (Microsoft.Android.Sdk.AssemblyResolution.targets): a resolve-only sibling of _ComputeFilesToPublishForRuntimeIdentifiers that stops after ResolveReferences (no ComputeFilesToPublish/ILLink/ILC/AOT) and returns the Android runtime-pack managed assemblies (Microsoft.Android.Runtime.*) that have no @(ReferencePath) counterpart (matched on Filename, so Mono.Android/Java.Interop aren't double-scanned; the .NET/BCL runtime pack is excluded).
• _ResolveRuntimeOnlyAssembliesForTrimmableTypeMap + _AddRuntimeOnlyAssembliesToTrimmableTypeMap (Microsoft.Android.Sdk.TypeMap.Trimmable.NativeAOT.targets): before _ResolveAssemblies spawns the per-RID ILC builds, run that target once via the MSBuild task for the first RID (managed metadata is RID-independent) and cache the result. The nested resolve is gated on Inputs/Outputs so it does not re-run on incremental no-op builds; an always-run loader reads the cached list into the generator's extra-framework input (_AndroidTrimmableTypeMapExtraFrameworkAssembly, plumbed through Microsoft.Android.Sdk.TypeMap.Trimmable.targets).
• ILC framework classification generalized: any runtime-pack managed assembly's per-assembly typemap DLL is treated as framework (IlcReference but not an unmanaged-entrypoint root), matching Mono.Android/Java.Interop, so its JCW native methods register via the runtime registerNatives path.

Test

Adds Build_WithTrimmableTypeMap_KeepsNativeAotRuntimeHostAcws, which opts into the trimmable typemap on NativeAOT and asserts classes.dex retains the UncaughtExceptionMarshaler runtime ACW.

Validation (local, -c Release)

• Build_WithTrimmableTypeMap_KeepsNativeAotRuntimeHostAcws passes.
• A NativeAOT + trimmable app build now generates _Microsoft.Android.Runtime.NativeAOT.TypeMap.dll (with the UncaughtExceptionMarshaler → scrc64…/UncaughtExceptionMarshaler entry + proxy), the JCW .java, and the acw-map entry.
• The runtime host typemap DLL is an IlcReference and not an UnmanagedEntryPointsAssembly (correct framework classification).
• No double-scanning of framework assemblies (one _Mono.Android.TypeMap.dll in the generator output).
• Incremental no-op builds skip the nested resolve (verified via binlog: _ResolveRuntimeOnlyAssembliesForTrimmableTypeMap up-to-date).

Notes

The trimmable typemap is opt-in on main (_AndroidTypeMapImplementation=trimmable); this fix is dormant otherwise. Extracted from the "make trimmable the NativeAOT default" work (#11822) so it can be reviewed and merged independently.

Issue references

Contributes to #10933 and #10793.

[Copilot]

Pull request overview

This PR fixes a NativeAOT + trimmable typemap startup crash by ensuring the typemap generator scans the runtime-only Microsoft.Android.Runtime.NativeAOT assembly (via a shipped reference assembly), so its runtime ACW(s) (notably UncaughtExceptionMarshaler) are kept by R8 and appear in classes.dex.

Changes:

• Add a new build test asserting UncaughtExceptionMarshaler is present in classes.dex for NativeAOT + trimmable typemap + R8.
• Extend _GenerateTrimmableTypeMap inputs to include an extra framework assembly item list for runtime-only assemblies.
• Produce + package a reference assembly for Microsoft.Android.Runtime.NativeAOT under tools/typemap-refs, and wire it into the NativeAOT trimmable typemap build.

Show a summary per file

File	Description
src/Xamarin.Android.Build.Tasks/Tests/Xamarin.Android.Build.Tests/TrimmableTypeMapBuildTests.cs	Adds regression test verifying the runtime ACW class is retained in classes.dex.
src/Xamarin.Android.Build.Tasks/Microsoft.Android.Sdk/targets/Microsoft.Android.Sdk.TypeMap.Trimmable.targets	Feeds additional runtime-only framework assemblies into typemap generation inputs.
src/Xamarin.Android.Build.Tasks/Microsoft.Android.Sdk/targets/Microsoft.Android.Sdk.TypeMap.Trimmable.NativeAOT.targets	Injects the NativeAOT runtime host ref assembly into typemap inputs; classifies its per-assembly typemap DLL as framework for ILC.
src/Microsoft.Android.Runtime.NativeAOT/Microsoft.Android.Runtime.NativeAOT.csproj	Enables ProduceReferenceAssembly and copies the ref assembly into the SDK pack’s typemap-refs staging area.
build-tools/create-packs/Microsoft.Android.Sdk.proj	Packages the NativeAOT runtime host reference assembly under tools/typemap-refs.

Copilot's findings

• Files reviewed: 5/5 changed files
• Comments generated: 2

[simonrozsival]

/review

[github-actions]

✅ Android PR Reviewer completed successfully!

[github-actions]

🤖 Code Review Summary

Verdict: ⚠️ Minor changes suggested — the MSBuild logic is well-designed and correct, and CI is fully green (42/42 dotnet-android checks passing on Azure DevOps). The two items below are non-blocking.

What this PR does well

• Solid root-cause fix. Runtime-only assemblies (the NativeAOT runtime host Microsoft.Android.Runtime.NativeAOT) were never in @(ReferencePath) for the RID-independent outer-build typemap generator. The auto-detect approach (a resolve-only sibling of _ComputeFilesToPublishForRuntimeIdentifiers that stops after ResolveReferences) is the right layer to solve this at, and generalizes to any runtime-only Java-peer assembly rather than hard-coding UncaughtExceptionMarshaler.
• Incrementality is carefully handled. The expensive nested resolve is gated on Inputs/Outputs, results are cached to runtime-only-assemblies.txt, and an always-run loader feeds the cached list into the generator's Inputs so the generator's own up-to-date check stays stable. The WriteOnlyWhenDifferent + FileWrites usage is correct.
• Re-entrancy is safe. The nested build sets _OuterIntermediateOutputPath, so _ResolveRuntimeOnlyAssembliesForTrimmableTypeMap/_AddRuntimeOnlyAssembliesToTrimmableTypeMap are correctly skipped inside it (no infinite recursion), mirroring the existing _ResolveAssemblies pattern.
• Good regression test. Build_WithTrimmableTypeMap_KeepsNativeAotRuntimeHostAcws asserts the UncaughtExceptionMarshaler ACW survives in classes.dex, correctly matching the CRC-hashed class-name suffix and gating on IgnoreUnsupportedConfiguration.

Findings (2 inline comments)

Severity	Category	Location
⚠️	Documentation	TrimmableTypeMapBuildTests.cs:96 — stale "reference assembly is shipped" comment contradicts the actual resolve-and-cache implementation.
💡	MSBuild targets	...TypeMap.Trimmable.NativeAOT.targets:37 — ILC framework classification includes the BCL runtime pack, unlike the resolve target which filters it out; harmless but asymmetric.

Neither blocks merge on correctness; the ⚠️ is a doc-accuracy fix worth making before merge. Since the trimmable typemap is opt-in on main, the risk surface is limited to _AndroidTypeMapImplementation=trimmable builds.

Reviewed against the dotnet/android MSBuild, testing, and repo-convention guidelines.

Generated by Android PR Reviewer for #11982 · 206.2 AIC · ⌖ 22.1 AIC · ⊞ 6.8K
_{Comment /review to run again}

[github-actions]

🤖 ⚠️ Documentation — This comment claims a "reference assembly for the host is now shipped in the SDK pack and fed to the generator," but that is not what the PR does. The final implementation auto-detects runtime-only assemblies at build time via the nested _ResolveRuntimeOnlyAssembliesForTypeMap resolve pass — the PR description itself states "No hard-coded assembly names, no per-assembly reference-assembly plumbing." This stale wording describes the abandoned ref-assembly approach (cf. the ...-refasm branch name) and will mislead anyone debugging this area later. Please reword it to describe the resolve-and-cache mechanism that the test actually exercises.

Rule: Remove stale comments (Postmortem #59)

[github-actions]

🤖 💡 MSBuild targets — This includes every runtime-pack managed DLL, including the .NET/BCL runtime pack (Microsoft.NETCore.App.Runtime.*), whereas the sibling _ResolveRuntimeOnlyAssembliesForTypeMap deliberately filters to Microsoft.Android.Runtime and excludes the BCL. It is harmless today — these names only feed the Remove on _TrimmableTypeMapUnmanagedEntryPointAssemblyNames, and BCL assemblies have no generated _*.TypeMap DLL to remove — but the asymmetry between the two targets is a maintenance trap and inflates the item group with hundreds of no-op BCL names. Consider matching the resolve target's StartsWith('Microsoft.Android.Runtime') filter here, or add a note explaining why the broader set is intentional.

Rule: Consistency with sibling targets