[MSBuild] Add retained typemap rule generation pipeline - #12827

Two issues in the staged typemap pipeline need fixing before it can be imported:

1. src/Xamarin.Android.Build.Tasks/Microsoft.Android.Sdk/targets/Microsoft.Android.Sdk.TypeMap.Proguard.targets:4,58-60,99-102: a trimmed R8 CoreCLR build with AndroidTypeMapImplementation=llvm-ir selects ExtractTypeMapKeysFromLlvmIr, but no task by that name is compiled into $(_MicrosoftAndroidBuildTasksAssembly) at this head. This path fails with MSB4036 instead of generating typemap keys. Please supply and test the extractor, or keep this branch inactive until it exists.

2. src/Xamarin.Android.Build.Tasks/Microsoft.Android.Sdk/targets/Microsoft.Android.Sdk.TypeMap.Proguard.targets:37-45: the NativeAOT enabled and disabled targets share proguard_project_references.cfg. On a true -> false change to _AndroidEnableTypemapR8Trimming, the ACW map and production _AndroidBuildPropertiesCache do not change (the cache does not record this flag). The fallback therefore skips as up-to-date and retains the narrower enabled-mode rules, defeating the opt-out. The new transition test masks this by using a synthetic cache that does record the flag. Please include the mode in fallback invalidation and test against production-equivalent cache contents.

I treated the product import and Dalvik/R8 wiring as intentionally deferred to Layer 4; these are defects in the staged pipeline itself.

⚠️ Needs Changes

Found 1 error and 1 warning in the new MSBuild pipeline: the complete ProguardConfigFiles override still executes an unused typemap generation path, and the member-rule target uses AfterTargets despite an available dependency extension point.

The producer metadata flow and focused incremental coverage are otherwise well designed. All 44 reported CI checks are green.

Generated by Android PR Reviewer for #12827 · copilot · gpt56 · 139.9 AIC · ⌖ 14 AIC · ⊞ 26.5K
_{Comment /review to run again}

✅ LGTM

Findings: 0 errors · 0 warnings · 1 suggestion

The pipeline correctly keeps exact producer outputs, runtime/mode state, custom-configuration gating, and Dalvik incrementality aligned. The new target is intentionally not imported by product targets in this layer, matching the PR description. All 44 reported checks are green. I left one non-blocking inline suggestion to add equivalent end-to-end CoreCLR/R8 coverage for the NativeAOT constructor-retention regression test.

Generated by Android PR Reviewer for #12827 · copilot · gpt56 · 150 AIC · ⌖ 11.2 AIC · ⊞ 26K
_{Comment /review to run again}

The previously reported extractor, NativeAOT member-retention, mode-invalidation, override and ordering concerns are resolved in current source. Product import and the stated Layer-4-dependent edits/tests remain accepted staging, not blockers. The complete eight-file review and separate integration pass identify one defect in the explicitly imported CoreCLR pipeline.

Keep linked-assembly extraction inactive when CoreCLR ILLink is disabled (TypeMap.Proguard.targets:125-128). With CoreCLR, PublishTrimmed=true, AndroidLinkTool=r8 and RunILLink=false, the new activation still invokes linked-DLL extraction, and the new inner-build inventory condition:85-91 still advertises typemap DLLs under $(IntermediateOutputPath)linked.

The repository's pinned SDK linker target excludes RunILLink=false, so on clean intermediates these linked DLLs are not produced. The existing CoreCLR publish adapter:317-320 can publish generated typemaps without linked copies; that does not create the linked paths advertised by this new inventory. Assembly extraction:33-44 opens those paths and reports missing files as XA4327, failing this non-linking configuration instead of leaving extraction inactive.

Exclude RunILLink=false from the CoreCLR activation and linked-inventory conditions and add a clean-intermediates case. Keep strict missing-input validation and exact linked producer contracts; do not substitute stale/pre-trim discovery. Do not apply this exclusion to NativeAOT, where ILC performs trimming.

This activation/producer mismatch is established by immutable source tracing, not an executed MSBuild reproduction. The observed CI failure in build1620088 is separately verified as a submodule checkout receive timeout before build/tests, and is not the reason for this source finding.

The disabled-ILLink activation/producer mismatch is fixed, and the earlier NativeAOT retention, mode-state, override and ordering findings remain resolved. The declared Layer-3 product-import deferral is accepted. The complete current eight-file review and separate integration/completeness pass identify one remaining incremental-contract blocker.

Track CoreCLR opt-out as a Dalvik input change (TypeMap.Proguard.targets:125-129). After an enabled trimmed CoreCLR/R8 build, a no-clean rebuild with _AndroidEnableTypemapR8Trimming=false skips the modern configuration/class/member targets before their dependencies. The effective configuration/file list changes, but the only mode-file writer is NativeAOT-only (lines 26-38). CoreCLR therefore does not update a policy sentinel or register an inactive-mode change with Dalvik.

The production properties cache does not record this flag, and the resolved-assembly hash ignores changed producer metadata. The real Dalvik inputs contain user @(ProguardConfiguration), not the internal selected configuration list; removing the modern input registrations does not invalidate the existing timestamp-based Dalvik stamp. When the remaining producer/consumer inputs retain their timestamps, R8 can skip and leave classes.dex from the previously selected keep configuration. The baseline CoreCLR configuration writer can also remain up to date; this finding does not assume that opt-out necessarily generates fresh legacy contents. This is pinned static source proof, not an executed rebuild or a claim that every flag flip skips R8.

Persist the effective CoreCLR policy mode even when modern extraction is inactive, include that state file in Dalvik inputs, and add a no-clean enabled/disabled/enabled consumer regression using unchanged linked DLLs and production-like cache/input contracts. Do not invoke inactive extraction/tasks or require Layer-4 product imports to fix this staged-pipeline contract.

Review-completeness correction: this gap already existed at the previously reviewed 2b4d048e0fb03db209b38399b6271c6eb7599908 head and was missed in the earlier complete review. It was not introduced by da429; the separate disabled-ILLink fix is valid.

The CoreCLR opt-out invalidation blocker is resolved: the policy-state writer now runs directly in the outer Dalvik dependency chain even when modern generation is inactive, updates a surviving consumer input with write-only-when-different semantics, and registers it for Clean. Re-enabling policy also invalidates class generation; disabled-linker/full-override paths avoid modern task invocation, and NativeAOT's shared-state behavior is preserved.

The complete eight-file/current-hunk review, separate producer-to-actual-Dalvik integration pass and final additional-blocker sweep found no remaining high-confidence source blocker. The declared Layer-3/Layer-4 staging and previously accepted policy decisions remain accepted. This is immutable source/contract review,not independently executed private R8,NDK,build or device validation.