Fix double GC relocation of overlapping byrefs

[jkoritzinsky]

Summary

Report each physical byref slot only once when scanning an explicit-layout byref-like value type. Overlapping fields currently cause the shared scanner to report the same slot multiple times, so a compacting GC can relocate an already-relocated interior pointer again.

This is the standalone GC fix requested in review of #133659. It is not debugger-specific: an ordinary by-value ref-struct argument reproduces the corruption on Windows x86 with DOTNET_GCStress=1.

• Track enclosing overlapping layouts and check earlier fields for the exact slot address, not the referenced object value. The check is stack-only and allocation-free; document its quadratic cost for rare overlapping layouts.
• Use the already-loaded value-type field handle rather than triggering type loading during GC enumeration.
• Add a targeted test to the existing ref-field tests, asserting all four expected values through both overlapping views. A non-inlined callee preserves the by-value argument across the JIT transition.

No debugger-specific or cDAC changes are included.

Validation

Windows x86 Checked runtime, Release libraries, at main base 6b1cd4b85c4b:

Check	Unfixed main	Fixed
Reviewer's repro, normal	1 2 3 4	1 2 3 4
Reviewer's repro, DOTNET_GCStress=1	0 3801155 7536732 2949173	1 2 3 4
Ref-field tests, normal	6/6 passed	6/6 passed
Ref-field tests, DOTNET_GCStress=1	New regression fails: expected 1, actual 6881312 (exit 101)	6/6 passed (exit 100)
Existing inline-array tests, normal / DOTNET_GCStress=1	Not run	11/11 passed in each mode (exit 100)

The regression assembly and CoreLib are byte-identical between the unfixed and fixed runs.

Build commands and results:

.\build.cmd clr+libs+host -a x86 -rc Checked -lc Release
# Passed; 0 errors, 3 existing MsQuic NU1902 warnings.
.\build.cmd clr.native+clr.corelib+clr.nativecorelib -c Checked -a x86
# Passed; 0 warnings, 0 errors.
.\src\tests\build.cmd -GenerateLayoutOnly x86 Checked /p:HostConfiguration=Debug
# Passed before and after the fix.
.\src\tests\build.cmd -Test Loader\classloader\RefFields\Validate.csproj x86 Checked /p:HostConfiguration=Debug
# Passed; 1 selected test project, 0 warnings/errors.
.\dotnet.cmd build -c Checked src\tests\Loader\classloader\InlineArray\InlineArrayValid.csproj /p:TargetArchitecture=x86 /p:BuildAsStandalone=true
# Passed; 0 warnings/errors.

Tests were run with CORE_ROOT set to artifacts\tests\coreclr\windows.x86.Checked\Tests\Core_Root, using & "$env:CORE_ROOT\corerun.exe" with Validate.dll and InlineArrayValid.dll from their respective output directories, once without DOTNET_GCStress and once with it set to 1. Other architectures/platforms were not run.

Resolves #135245

Note

This implementation and PR description were prepared with GitHub Copilot.

[azure-pipelines]

Azure Pipelines:
Successfully started running 3 pipeline(s).
13 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: @agocke
See info in area-owners.md if you want to be subscribed.

[Copilot]

🔵 Needs a closer look

Safety-critical GC scanning changes need runtime-owner review of scanner invariants and platform coverage beyond Windows x86.

No open issues 🎉

What changed in this PR

Fixes double GC relocation by reporting each physical byref slot once in overlapping ref-struct layouts, independently of debugger functionality.

Changes:

• Deduplicates overlapping slots using allocation-free traversal.
• Uses already-loaded field type handles during GC scanning.
• Adds regression coverage for nested overlaps and shared backing arrays.

Validation reported in the PR covers Windows x86; no builds or tests were run during this review.

File	Description
src/​tests/​Loader/​classloader/​RefFields/​Validate.cs	Tests overlapping and nested byrefs through by-value arguments.
src/​coreclr/​vm/​siginfo.cpp	Prevents duplicate slot reporting and avoids type loading during scanning.

_{Review effort: Balanced}

[jkoritzinsky]

/ba-g failures unrelated