[ASDAlexander77]

Follow-up to #639, covering the two cases it left out of holdGlobalArguments.

Shared handles: fixed

run(bare), where bare is a Shared<T> global and run gives bare a new value, freed the block the parameter still read. This happened under rc and own alike: own counts handles with rc's routines, and its #639 rule (rejectGlobalArgumentsDropped) skips handles.

Holding every handle argument would change what Shared.count reports inside the callee. Spec 23.1 defines that count as the number of handles, and a parameter is a borrow, not a handle. So the hold is narrowed instead:

• A handle is held only across a call that may assign the global it was read from: directly, in a callee (a transitive per-function set of assigned globals), or through a call that names no function (counted as possibly assigning anything). A store into a place inside the global (a field, an element, s.value) counts as assigning it.
• So no existing Shared.count expectation changes: every own_shared_* test passes unmodified. Inside a call that does assign the global, the hold is the count that keeps the block alive once the global has moved on. The new test asserts it is then the only count.
• Under own, only a handle read straight out of a global is held. A handle read out of a global's field is still a borrow of that field, and own rejects it with the same error as on main.
• Under rc, a value read through a global handle (run(G.value)) is now also recognised as global-rooted and held, like a field read.

Calls in a try body (ts.Invoke): did not reproduce as described

This pass runs before the affine lowering, and it is the affine lowering that turns a call inside a try body into a ts.Invoke. So those calls were already held: Retain before the call, and the Release ends up first in the Invoke's normal successor (checked in --emit=mlir-affine). The remaining gap is the unwind edge: when the callee throws, that one count leaks. It can't be observed for a plain class, and --verify-ownership does not see unwind edges. Fixing it needs a per-call release on the exception path, which is a new mechanism. It is left open, and the comment that said try-body calls were skipped is corrected.

Tests

• own/own_shared_global_argument.ts, in the shared_models foreach (gc/none/rc/own) and the rc/own --verify-ownership list. It covers a direct assignment, an assignment through a callee, a call in a try body, and checks that a call which doesn't assign takes no count. On main it fails under rc and own.
• tests/00shared_global_field_argument.ts, in the default lists and TSLANG_CORPUS (rc + none): a handle read out of a global's field. On main it fails under rc.
• Teeth: skipping the handle hold fails test-compile-rc-own_shared_global_argument, test-compile-own-own_shared_global_argument and test-compile-rc-corpus-00shared-global-field-argument (and their jit twins). Without the switch they pass.
• 00shared_global_field_argument.ts also pins run(G.value): a value read through a global handle, given to a call that replaces it (fails on main under rc).
• Windows ctest -C Release: 4930/4930. WSL Linux full run: 4912/4912 (before the G.value case was added; that test-only commit was re-run on Windows only).

🤖 Generated with Claude Code