Remove CoreCLR Reflection.Emit's ceefilegen dependency - #134966

CoreCLR's in-memory Reflection.Emit implementation uses PE-style sections and relocation records just to retain method IL and FieldRVA bytes. Store those bytes directly in loader-owned memory and reuse the existing token-to-address map, so shared and static CoreCLR no longer need to link ceefgen.

Changes

• Allocate emitted IL and aligned, zero-initialized field data on the module's loader allocator. Preserve IL/EH encoding and collectible field-data lifetime tracking.
• Use MethodDef tokens as non-address values in the dynamic module's metadata RVA column, retaining the nonzero-body invariant. Resolve field data using the existing FIELD_OFFSET_DYNAMIC_RVA sentinel and token map, including reflection's fast field-access path.
• Remove generator ownership, PE section allocation, and native/managed token-relocation bookkeeping. Preserve ILAsm's separate ceefgen_nohost dependency and PE-writing implementation.
• Add emission, field-content/alignment, and diagnostic coverage. Update the canonical cDAC descriptor descriptions and generated documentation.

Existing token-aware diagnostic paths resolve the new storage. Raw ISOSDacInterface.GetILForModule support for Reflection.Emit remains unchanged: it is unsupported by the native DAC.

Validation

Windows x64 Release and Checked builds passed. Local coverage includes:

• All 2,968 Reflection.Emit, ILGeneration, and Lightweight library tests.
• 408 cDAC Loader/MethodTable tests, including mock 32/64-bit and little/big-endian layouts and legacy-facing entry points.
• 18 runtime reflection wrappers, collectible byref/span lifetime coverage, profiler ModuleLoad and ReJIT tests, and 24 existing EnC tests.
• The 12 new behavior-preservation cases on both the frozen baseline and changed Release runtime.
• Documentation generation/drift checks and actual Checked/Release Ninja link inputs confirming that shared/static CoreCLR exclude ceefilegen while ILAsm retains it.

Other runtime architectures and metadata-updater-disabled builds were not run. The pre-existing commented-out EnC TestAddFieldRVA case was left unchanged and was not included in the EnC count.

Local measurements and trade-off

Dependency-free, ad hoc Windows x64 Release harness; A/B/B/A process order, 18 measurements per operation per host, and tiered compilation disabled for both. These are shared-machine measurements, not BenchmarkDotNet results.

The slowdown below is limited to the non-intrinsic runtime fallback path for FieldRVA data defined by System.Reflection.Emit. It is not a loss of intrinsic expansion: these benchmark cases use the fallback in both the baseline and changed builds.

Module-creation medians were approximately 0.3%-4.1% lower across four workloads. Process-private-memory growth while retaining emitted modules was approximately 24%-39% lower; this is not a pure native-allocation metric.

The slow benchmark supplies field handles obtained through reflection and reuses an existing array. Its fallback now resolves emitted field data through a locked token-map lookup instead of unlocked section-offset translation. Follow-up disassembly confirmed identical managed fallback instruction sequences in both builds.

Intrinsic-friendly ldtoken/CreateSpan and constant-size newarr/dup/ldtoken/InitializeArray patterns still expand to direct data access in both builds, with no per-invocation token lookup; the address is resolved at JIT time. Ordinary PE-backed fields also retain their existing lookup path. Follow-up controls for those cases were approximately unchanged, so the slowdown percentages above do not apply to them.