Add regression test: #12796, DefaultValue null on array type record field

[T-Gro]

Fixes #12796

Adds a regression test verifying that [<DefaultValue(null)>] on a record field of type A[] compiles successfully without internal error FS0192.

[github-actions]

⚠️ Release notes required, but author opted out

Warning

Author opted out of release notes, check is disabled for this pull request.
cc @dotnet/fsharp-team-msft

[T-Gro]

Big NO, the issue really says:

[<DefaultValue([||] : A[])>]

[github-actions]

Commit pushed: 6d1be63

Generated by Regression PR Shepherd

[github-actions]

Commit pushed: 7128e25

Generated by Regression PR Shepherd

[github-actions]

🤖 This is an automated response from Regression PR Shepherd.

This regression test proves the bug in #12796 still exists on Desktop .NET Framework. Closing this PR is recommended.

The bug: [<DefaultValue([||] : A[])>] on a record field of type A[] triggers an internal compiler error FS0192: encodeCustomAttrElemType: unrecognized custom element type. The issue was opened in 2022, and was labeled AI-thinks-issue-fixed after the bot detected SynPat.Record and SynPat.QuoteExpr fixes — but that was a different issue; this bug is specifically about attribute encoding for array-typed DefaultValue arguments.

The test: tests/FSharp.Compiler.ComponentTests/Language/AttributeCheckingTests.fs typechecks type B = { [<DefaultValue([||] : A[])>] BField: A[] } and expects success.

The failure: Build WindowsNoRealsig_testDesktop fails — the test fails only on Desktop .NET Framework, not on CoreCLR/Linux. The FS0192 error is triggered during attribute argument encoding, which happens during type checking on Desktop.

The AI-thinks-issue-fixed label has been removed from #12796.

cc @T-Gro @abonie

Generated by Regression PR Shepherd · ◷

[T-Gro]

I don't get why we would have problems encoding empty arrays of any type? It does not matter if we cannot store instances of that type, since it is empty. If we can store a string array, an empty string array, why also not an arbitrary empty array?

[T-Gro]

@copilot :

See my comment above. Research implementation changes needed to support empty arrays as default values, and develop using TDD (via ComponentTests. Since this is IL specific, make sure to use "compileAndRun" and also use the helper for "ilverify" of the compiled code in that test.

[Copilot]

@copilot :

See my comment above. Research implementation changes needed to support empty arrays as default values, and develop using TDD (via ComponentTests. Since this is IL speci...

Implementation complete in commit d2be16e.

Implementation changes:

• IlxGen.fs (GenAttribArg): Detects array expressions where the element type is an unencodable reference type (ILType.Boxed that isn't String/Object/Type). For empty arrays, substitutes System.Object as the element type (safe since no elements need encoding). For non-empty arrays, produces FS3885 diagnostic instead of the internal error.
• il.fs: Added 0x51 (SERIALIZATION_TYPE_TAGGED_OBJECT) decoder support so the compiler can read back its own object[]-typed attribute blobs.

Tests (all using compile → verifyPEFileWithSystemDlls → run pipeline):

• Empty array of class type ([<DefaultValue([||] : A[])>] — the exact issue repro)
• Empty array of record type
• Empty array of enum type
• Empty array of primitive type (int)
• Negative: non-empty array of class type → FS3885

All 97 AttributeCheckingTests pass.