[ASDAlexander77]

Two class-field bugs with no interface involved, found while landing #612.

1. fn = () => 0 then a capturing closure

class W2 { fn = () => 0; constructor(k: int) { this.fn = () => k; } }
function main() { const w = new W2(9); print(w.fn()); }

warned "losing this reference" and crashed (0x80000003). The field's type was inferred from the initializer as a plain !ts.func, which has no room for a closure's context. A field whose function type is inferred is now stored as a hybrid function, as an annotated fn: () => int field and a let (adjustLocalVariableType) already are.

2. fn = this.m

class W3 { v = 4; m(): int { return this.v; } fn = this.m; }
function main() { const w = new W3(); print(w.fn()); }

failed with "not a function to call". The real pass types the field as the bound method, but the discovery pass lays out the fields before it declares the methods, so the field was any there and the call failed before the real pass ran. In a discovery pass such a field now holds its place and is typed again once the members are declared (mlirGenClassRetypeFields). Code emitted against the placeholder is thrown away with the pass.

The field type computation moved into mlirGenClassDataFieldType so that both places use it.

Tests

00class_function_typed_field_inferred.ts covers both forms, reassignment (including a closure replacing the bound method) and the interface path from #612. It runs under gc AOT/JIT and in the rc/none corpus. Before the fix it failed in all 6 modes; it now passes in all of them. Full ctest -C Release passes: 4736/4736.

These also work: a class used before its declaration, a class local to a function, and fn = this.m naming an inherited method. Under -mm=rc --verify-ownership the test produces no diagnostics.

🤖 Generated with Claude Code