[ASDAlexander77]

Why it was refused

Under rc and own an object's count is the word before it. A @cpp class's reference is a C++ pointer, and the word before a C++ object belongs to the C++ heap (or a neighbour), so a retain or release would write into C++ memory.

Change

No model counts a @cpp reference now:

• ClassStorageType carries a cpp flag, set right where the class type is made (MLIRGenClasses.cpp), before anything of the type is generated.
• MLIRTypeHelper::ownsHeapMemory answers no for it. MLIRGen, the ownership passes and the release/retain routines all ask that, so copies into locals, fields, arrays and arguments retain and release nothing, and a holder's destroy skips a @cpp field.
• The "needs the gc or none memory model" error is removed.

The object stays C++'s to free. An object of a class implementing C++ methods made with new is never freed by tslang under rc/own (documented in docs/cpp-classes.md, new "Memory models" section).

Tests

• 00cppvtable_class.ts, 00cpp_slot.ts, 00cpp_overloads.ts join the rc/none corpus (and the ownership verifier over it), and run under own (AOT + JIT).
• cppvtable/memory_model.ts compiles under rc, own and none.
• test-jit-cppvtable-abi also runs under rc and own, plus --verify-ownership. abi_host.cpp's new heap object sits right after a guard word of C++'s own; the guard is checked while tslang holds copies (field, array, local) and after, then C++ deletes the object.
• Teeth: with the exemption turned off, the rc run fails "nothing written in the word before a C++ object" and the own runs fail to compile.
• Windows Release ctest: 4857/4857 passed. The Itanium (Linux) half got the same change but was only run by CI.

🤖 Generated with Claude Code