[XABT] Fix stale shrunk assemblies in the llvm-ir typemap path by BrzVlad · Pull Request #12342 · dotnet/android

BrzVlad · 2026-08-11T11:26:30Z

Summary

Fix stale shrunk assemblies in the llvm-ir typemap path and restore the Release MonoVM APK smoke test disabled in #12267.

_RemoveRegisterAttribute copies @(_ResolvedAssemblies) into @(_ShrunkAssemblies). Packaging reads the shrunk copies, while mono-aot-cross compiles the resolved assemblies. The two sets could diverge because:

Ordering: _RemoveRegisterAttribute ran before _GenerateJavaStubs, which invokes RewriteMarshalMethods and rewrites @(_ResolvedAssemblies) in place, removing the cb_* JNI callback fields. _PostTrimmingPipeline can also rewrite the linked assemblies after ILLink.
Invalidation: $(_AndroidLinkFlag) tracks ILLink, but it does not observe those later in-place rewrites. Tracking only the link flag allowed the shrunk copies to remain up to date while the resolved assemblies changed.

Changes

Make _RemoveRegisterAttribute depend on _GenerateJavaStubs so it copies the final rewritten assemblies.
Preserve $(_AndroidLinkFlag) as an input for ILLink invalidation and additionally track:
- @(_ResolvedAssemblies) for post-link in-place rewrites.
- $(_ResolvedUserAssembliesHashFile) for changes to the resolved assembly set.
- $(_AndroidBuildPropertiesCache) and @(_AndroidMSBuildAllProjects) for relevant build/project changes.
Add $(_RemoveRegisterFlag) to @(FileWrites) so clean/incremental-clean behavior tracks the stamp.
Re-enable the Mono.Android.NET_Tests-Mono Release MonoVM instrumentation lane that [main] Update dependencies from dotnet/dotnet, microsoft/testfx #12267 temporarily disabled.

This aligns the llvm-ir path with the incremental-input model used by _RemoveRegisterAttributeCoreClr in Microsoft.Android.Sdk.TypeMap.Trimmable.targets, while retaining the existing ILLink sentinel.

Failure mechanism

In #12267, the packaged Mono.Android.dll retained Android.Widget.AbsListView::cb_getAdapter while the AOT input did not. This shifted the metadata Field table rows for Android.Runtime.AndroidTypeManager. AndroidTypeManager..cctor then resolved an AOT SFLDA patch to the instance field jniAddNativeMethodRegistrationAttributePresent instead of the static prevent_delegate_gc_lock, causing Mono to abort during startup before any tests ran:

Assertion at mono/metadata/object.c:3109,
condition `field->type->attrs & FIELD_ATTRIBUTE_STATIC' not met

Mono's AOT staleness check did not catch the mismatch because it compares assembly MVIDs, and Cecil preserves the MVID across these rewrites.

Validation

Validated on a Pixel 7a (arm64, API 36) with Mono.Android.NET-Tests, Release + -p:UseMonoRuntime=true, from clean obj directories:

Before: deterministic crash in 5 runs; linked assembly had 1767 field rows while the shrunk copy had 1768.
After: 281 tests passed with 0 failures in 2 runs; linked and shrunk assemblies both had 1767 field rows.
Control: reverting the target ordering/input changes restored both the crash and the 1767/1768 mismatch.
Azure DevOps: all 44 PR checks pass, including the restored Mono APK smoke lane.

Fixes dotnet/runtime#131760

`_RemoveRegisterAttribute` copies `@(_ResolvedAssemblies)` into `@(_ShrunkAssemblies)`. Packaging reads the shrunk copies, while `mono-aot-cross` compiles the resolved assemblies. Two defects let those two sets diverge: * Ordering: the target ran before `_GenerateJavaStubs`, which invokes `RewriteMarshalMethods` and rewrites `@(_ResolvedAssemblies)` in place, removing the `cb_*` JNI callback fields. `_PostTrimmingPipeline` rewrites them again `AfterTargets="ILLink"`. * Invalidation: `Inputs` tracked `$(_AndroidLinkFlag)`, which is only touched by ILLink and so never changes when those post-link rewrites modify the assemblies. The result is an APK whose AOT images were built from assemblies with a different metadata `Field` table than the assemblies actually packaged. AOT encodes a static field reference as a class reference plus a `Field` table row (`encode_field_info` / `decode_field_info`), so one added or removed field shifts every later row. In dotnet#12267 the packaged `Mono.Android.dll` retained `Android.Widget.AbsListView::cb_getAdapter` while the AOT input did not, moving `Android.Runtime.AndroidTypeManager` from rows 984-987 to 985-988. `AndroidTypeManager..cctor` then resolved its `SFLDA` patch to row 985 - the instance field `jniAddNativeMethodRegistrationAttributePresent` instead of the static `prevent_delegate_gc_lock` - and Mono aborted during startup, before any test ran: * Assertion at mono/metadata/object.c:3109, condition `field->type->attrs & FIELD_ATTRIBUTE_STATIC' not met Mono cannot detect the mismatch on its own: its AOT staleness check compares the assembly MVID, and Cecil preserves the MVID across rewrites, so both copies look identical to it. `_RemoveRegisterAttributeCoreClr` in Microsoft.Android.Sdk.TypeMap.Trimmable.targets already keys `Inputs` off `@(_ResolvedAssemblies)`. This brings the llvm-ir path in line, which is why only the Mono leg regressed while CoreCLR passed. Validated on a Pixel 7a (arm64, API 36) with the Mono.Android.NET-Tests instrumentation, Release + `-p:UseMonoRuntime=true`, each run from a clean `obj`: * before: crash (5 runs, deterministic), linked/ 1767 rows vs linked/shrunk/ 1768 rows * after: 281 tests, 0 failed (2 runs), linked/ and shrunk/ both 1767 rows * control: reverting only these two attributes on the same built SDK restores both the crash and the 1767/1768 mismatch The shrunk copies were already rewritten on every build before this change, so tracking `@(_ResolvedAssemblies)` does not regress the incremental work in dotnet#12229. Fixes dotnet/runtime#131760 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 56a10519-c924-45fd-a767-83cd4e5f947a

Copilot

Pull request overview

This PR fixes a build-order and incremental-build invalidation issue in the llvm-ir typemap pipeline that could cause “shrunk” assemblies (used for packaging) to diverge from the assemblies compiled by mono-aot-cross, leading to runtime AOT failures due to mismatched metadata field-table row indices.

Changes:

Move _RemoveRegisterAttribute behind _GenerateJavaStubs so it copies from the final, post-rewrite linked assemblies.
Improve incremental invalidation by keying _RemoveRegisterAttribute off @(_ResolvedAssemblies) (plus project/build cache inputs) instead of $(_AndroidLinkFlag).
Track the stamp file in @(FileWrites) so it is cleaned reliably.

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

Restore the Release MonoVM instrumentation lane now that the llvm-ir typemap path keeps packaged shrunk assemblies in sync with the AOT inputs. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Keep the existing link flag input while also tracking final rewritten assemblies and the resolved assembly set hash. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

jonathanpeppers · 2026-08-11T22:00:11Z

/review

github-actions · 2026-08-11T22:00:54Z

❌ Android PR Reviewer failed. Please review the logs for details.

jonathanpeppers · 2026-08-11T22:06:15Z

@dalexsoto review

dalexsoto

The target ordering and invalidation changes keep packaged shrunk assemblies aligned with the Mono AOT inputs, and the stamp/output cleanup is coherent. No blocking issues found.

Copilot AI lite review requested due to automatic review settings August 11, 2026 11:26

BrzVlad requested review from jonathanpeppers and simonrozsival as code owners August 11, 2026 11:26

Copilot started reviewing on behalf of BrzVlad August 11, 2026 11:27 View session

BrzVlad mentioned this pull request Aug 11, 2026

[Mono] [Android] Assertion in mono_static_field_get_addr during app startup after dotnet/dotnet update dotnet/runtime#131760

Closed

Copilot AI reviewed Aug 11, 2026

View reviewed changes

Comment thread ...droid.Build.Tasks/Microsoft.Android.Sdk/targets/Microsoft.Android.Sdk.TypeMap.LlvmIr.targets Outdated

[ci] Re-enable Mono APK smoke test

164057c

Restore the Release MonoVM instrumentation lane now that the llvm-ir typemap path keeps packaged shrunk assemblies in sync with the AOT inputs. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

jonathanpeppers reviewed Aug 11, 2026

View reviewed changes

Comment thread build-tools/automation/yaml-templates/stage-package-tests.yaml

[XABT] Preserve ILLink invalidation for shrunk assemblies

692ec45

Keep the existing link flag input while also tracking final rewritten assemblies and the resolved assembly set hash. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

jonathanpeppers approved these changes Aug 11, 2026

View reviewed changes

dalexsoto approved these changes Aug 11, 2026

View reviewed changes

jonathanpeppers merged commit 6308658 into dotnet:main Aug 11, 2026
44 checks passed

Navigation Menu

Uh oh!

[XABT] Fix stale shrunk assemblies in the llvm-ir typemap path - #12342

[XABT] Fix stale shrunk assemblies in the llvm-ir typemap path#12342
jonathanpeppers merged 3 commits into
dotnet:mainfrom
BrzVlad:fix-mono-testing2

BrzVlad commented Aug 11, 2026 •

edited by jonathanpeppers

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

jonathanpeppers commented Aug 11, 2026

Uh oh!

github-actions Bot commented Aug 11, 2026 •

edited

Loading

Uh oh!

jonathanpeppers commented Aug 11, 2026

Uh oh!

dalexsoto left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

BrzVlad commented Aug 11, 2026 • edited by jonathanpeppers Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Changes

Failure mechanism

Validation

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

jonathanpeppers commented Aug 11, 2026

Uh oh!

github-actions Bot commented Aug 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jonathanpeppers commented Aug 11, 2026

Uh oh!

dalexsoto left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

BrzVlad commented Aug 11, 2026 •

edited by jonathanpeppers

Loading

github-actions Bot commented Aug 11, 2026 •

edited

Loading