Investigate additional NativeAOT static initialization and rooting size opportunities #132595

Summary

Follow-up to #132550 and #132576. A review of all 15 NativeAOT Sizeoscope reports linked from #132576 (comment) found additional cases where static initialization, canonical generic sharing, reflection metadata, or shared representations retain unnecessary code and data.

Priority order: standard/base applications first, representative real-world applications second, and explicitly minimal configurations as supporting evidence. The strongest near-term candidates are Unix ConsolePal, ASP.NET HostString, typed-header generic sharing, and hosting telemetry method normalization. The largest longer-term opportunity is NativeAOT AppContext initialization.

Candidate overview

Direct candidates

• Defer Unix ConsolePal terminal metadata. Move Lazy<TerminalFormatStrings> into a nested holder. Plain Console.WriteLine currently roots TermInfo.DatabaseFactory.ReadActiveDatabase without using terminal formatting.
• Defer ASP.NET HostString implementation caches. Move s_safeHostStringChars and s_idnMapping into non-reflected nested holders, independently where useful. The current constructor is rooted directly through reflection of the private SearchValues<char> field.
• Replace HostingTelemetryHelpers.KnownHttpMethods. Reuse HttpMethods canonical comparisons or an equivalent allocation-free helper instead of constructing a FrozenDictionary.
• Rewrite PathUtils table initialization without LINQ. Both validation tables are used in the server workloads, so holder separation may be neutral; directly construct the small platform-specific sets and measure.
• Separate independent Task static groups. Give TaskFactory (~120 B), the completed-task singleton (~362 B), and the execution-context callback (~80–126 B) independent holders.

Prototypes requiring measurement

• Specialize HeaderDictionaryTypeExtensions.GetList<T> for NativeAOT. Canonical System.__Canon sharing causes RequestHeaders.Accept to retain the Set-Cookie parser. Use closed typed parsing entry points or an inlinable type dispatcher.
• Separate Linux NetworkChange address and availability state. Address-change subscription currently roots availability-only dictionaries, event arguments, callbacks, timer machinery, and handlers.
• Separate UriHelper whitespace and search tables. Uri.ResolveHelper needs the whitespace array but triggers five SearchValues<char> initializations. Retained escaping methods may still root a subset.
• Avoid routing HMAC long-key normalization through every SHA implementation. HMACSHA256 -> HMACCommon.ChangeKeyImpl -> SHA3_512.HashData roots SHA3 support initialization. Investigate provider-based hashing using the selected algorithm identifier.
• Investigate TimeZoneInfo.CreateUtcTimeZone. Many retained methods reference UTC state, so simplifying UTC construction/localization is more promising than only moving the field.
• Evaluate compact HttpCharacters validation tables. Seven shared SearchValues produce 3.7–4.3 KB closures. Full Kestrel uses most operations, so holders alone are unlikely to help; compact ASCII tests or RVA bitsets require performance validation.

Architectural investigations

• Avoid eager NativeAOT AppContext materialization. A holder is insufficient because AppContext.GetData is retained. Investigate on-demand native runtime-knob lookup instead of eagerly decoding all keys and values into Dictionary<string, object?>.
• Separate KnownHeaders name recognition from object materialization. One reflected header getter initializes every KnownHeader and parser. TryGetKnownHeader references every field, so per-field holders alone cannot remove the closure.
• Reduce eager logging delegate construction. Kestrel, routing, Data Protection, gRPC, and health checks contain many independent LoggerMessage.Define closures. Evaluate source-generated logging or per-event implementations.

Ecosystem follow-ups

• Avalonia MicroCom registration: the Windows module initializer eagerly registers independent vtables, with individual closures of roughly 3–15 KB. Investigate per-interface registration holders or registration from actual consumers.
• Avalonia reflection metadata: compiler-generated static-constructor lambdas and broad property-registration graphs are retained. Investigate excluding unnecessary implementation methods and narrowing reflected control/property roots.

Coverage

Not currently actionable as holder-only changes

The NativeAOT reflection type-unifier tables are already independently isolated. RuntimeAssemblyInfo, runtime cast/interface caches, processor and array-pool state, calendar/leap-second state, generated JSON/protobuf metadata, and both BCryptAlgorithmCache paths are genuinely used in the examined workloads. Encoding singletons and comparer wrappers are retained by shared runtime selection logic. TextWriter.Null is constrained by its public static field shape.

The reports also contain the Number roots already addressed by #132550 and the CompareInfo/EncodingTable roots addressed by #132576.