[jtschuster]

Warning

This is a totally AI-generated, unreviewed implementation draft.
The implementation and this PR description were generated by GitHub Copilot and have not received human code review. Passing local tests is not a substitute for review or proof of conformance to the specification. This is not ready to merge.

Intent

This should implement the partitioned WebAssembly ReadyToRun (R2R) module specification discussed in #135544, including physical auxiliary modules and MIBC-guided hot/cold placement. It includes the earlier delayed-activation prototype from #136498.

The goal is to enable useful size and performance measurements once we have representative, execution-derived profiles. No real Havit startup profile has been collected yet. The current MIBC measurements use synthetic method lists and do not establish realistic startup size, startup performance, or a production-quality hot set.

The design and implementation notes describe the intended contracts and the limits of the existing evidence.

Implementation

• Emit a primary module, an auxiliary entrypoint-mapping module, and a cold-body module with its owned fixup, GC/unwind, and exception-handling data.
• Keep unavailable methods callable through interpreter fallback; activate mapping and body modules independently, in either order, with activation status published last.
• Integrate loading and activation with the browser host and corerun, and integrate publication, compression, and fingerprinting with the WebAssembly SDK.
• Use the union of explicitly supplied MIBC method membership for placement, including canonical generic matching. Mandatory-primary methods remain resident. An explicit empty/unmatched profile differs from having no profile; no-profile publishing retains the heuristic.
• Avoid implicitly enabling partial compilation when an auxiliary publish supplies a profile, so the profile controls placement rather than removing the remaining compiled code. Include the generic entrypoint-offset repair exposed by the added coverage.

The implemented boundary is single-threaded CoreCLR Browser-WASM, composite R2R, with retained IL. This is not non-composite, multithreaded, or stripped-IL support.

Using profiles

With matching updated compiler and SDK build targets, add the existing item to the application:

<ItemGroup>
  <PublishReadyToRunPgoFiles Include="startup.mibc" />
</ItemGroup>

dotnet publish \
  -p:PublishReadyToRun=true \
  -p:PublishReadyToRunComposite=true \
  -p:WasmEnableAuxiliaryR2RModules=true

Crossgen2 also accepts --wasm-auxiliary-modules --mibc startup.mibc directly. This consumes existing MIBC data; it does not implement browser profile collection. Selection is by method membership, not a CPU-weight threshold.

Measurements so far

The committed documentation contains the earlier nine-app, no-profile comparison and a separate matched Havit MIBC size-sensitivity comparison. These measurements predate the merge of current main and were not regenerated after that merge. They must not be interpreted as measured startup PGO results.

For Havit, both sides used the same updated compiler and retained 113,991 managed function/funclet slots. Brotli-compressed R2R module sizes in KiB:

Module	No MIBC heuristic	Synthetic four-method MIBC
Primary	1111.5	378.3
Entrypoint map	699.5	750.4
Cold body	3286.9	3822.6
Combined	5097.8	4951.4

This is an almost-all-cold configuration: the synthetic profile contains only String.Concat(string, string), String.IsNullOrEmpty(string), Int32.ToString(), and String.StartsWith(string, StringComparison). It demonstrates placement and size sensitivity, not a suitable startup hot set. These are module-file sizes, not whole-app completed network transfers.

The earlier heuristic comparison showed smaller primary modules but increased combined R2R payload and regressions in initial interactions and some steady-state throughput. This PR does not claim an overall performance improvement. Long synchronous managed calls can block JavaScript-side activation; representative profiles and activation-aware measurements are still needed.

Exact raw/gzip/Brotli inventories:

• Nine-app heuristic comparison.
• Six Havit profile variants.

Local validation

The merged feature was validated at e01332100e8, which incorporates upstream main at 82b3c9469e092824425186a37ace7b6ab6bc89b0.

./build.sh clr+libs+host -os browser -arch wasm -c Release /p:RuntimeFlavor=CoreCLR

Passed: Browser-WASM Release runtime, libraries, and host build, zero warnings/errors.

dotnet build \
  src/coreclr/tools/aot/ILCompiler.ReadyToRun.Tests/ILCompiler.ReadyToRun.Tests.csproj \
  --no-restore -c Release /t:Build,Test \
  /p:TargetOS=browser /p:TargetArchitecture=wasm \
  /p:CoreCLRConfiguration=Release /p:LibrariesConfiguration=Release

Passed: 233 tests, 39 skipped, zero failures, zero build warnings/errors. The six new MIBC selection cases failed before the selector change and passed afterward.

npm --prefix src/native run test:auxiliary-modules

Passed: 17 tests, zero failures.

src/tests/build.sh -browser -release -generatelayoutonly -skiprestorepackages /p:LibrariesConfiguration=Release

Passed: refreshed Browser-WASM Core_Root. The installed Crossgen2 hash matched the newly built compiler, and the runtime/CoreLib timestamps were checked.

The following wrapper was run for each of PartitionedR2RActivation_Eager, PartitionedR2RActivation_EntryPointsFirst, PartitionedR2RActivation_ModuleFirst, PartitionedR2RActivation_AllHot, and PartitionedR2RActivation_AllCold:

bash artifacts/tests/coreclr/browser.wasm.Release/readytorun/wasm/DelayedR2RActivation/<variant>/<variant>.sh \
  -coreroot "$PWD/artifacts/tests/coreclr/browser.wasm.Release/Tests/Core_Root"

Passed: all five physical runtime variants, wrapper exit 0 and actual managed exit 100. Each variant's cached IL-CG2 directory was removed before execution, and a fresh Crossgen2 invocation was verified in every log.

Additional retained evidence from the pre-merge checkout at 53380e2f8bb:

• Six real Havit publishes and six browser smokes passed, with five rendered UI navigations each, both activation cells active, and a real native mapping-activation notification.
• Six direct MSBuild profile-flow cases passed after the SDK fix; the auxiliary/profile case failed before it.
• A CoreLib-inclusive native run with an explicit unmatched MIBC passed activation and 200 exceptions, 400 finally executions, and 400 forced collections with live roots.
• Earlier V1 validation covered both activation orders, idempotence, nonzero bases, all-hot/all-cold/empty boundaries, and five physical runtime variants.

The extended Wasm.Build.Tests xUnit fixture was not built or run because its assets/test SDK were not prepared. Direct target probes and real publishes were used for the SDK change instead. Publishing used an isolated SDK package copy with the changed source target overlaid; archived runtime/SDK packages and earlier measurement outputs were preserved.

Post-merge build/test logs are under artifacts/aux-pr-merged-* in the feature checkout. Earlier profile/measurement commands and raw results are indexed by artifacts/mibc-selection-validation.json in the feature checkout and artifacts/mibc-hotcold-full-f8b896-20261028/report.json in the macro-benchmarks checkout. These local artifacts are not attached to this PR.

Draft status and remaining work

Human review of specification conformance, ownership/activation safety, code generation, SDK integration, and tests is required. Real startup/interaction profiles and matched size/performance measurements are also required before drawing conclusions about the tradeoff.

This branch includes the phase-1 prototype ancestry. Upstream main was incorporated with explicit merge commit e01332100e8; both ReadyToRun format documentation conflict hunks were resolved in favor of main, and that file now matches main exactly. The merged runtime/compiler and physical activation tests passed as described above.

Note

This implementation and PR description are entirely AI/Copilot-generated and have not been human-reviewed.

[azure-pipelines]

Azure Pipelines:
Successfully started running 4 pipeline(s).
12 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/crossgen-contrib
See info in area-owners.md if you want to be subscribed.

[dotnet-policy-service]

Tagging subscribers to 'arch-wasm': @lewing, @pavelsavara
See info in area-owners.md if you want to be subscribed.

[jkotas]

stride-indexed encoding

This will be gone once you resolve conflicts with main