[ASDAlexander77]

This closes the gap left open in #643. holdGlobalArguments retains a global read given to a call that may overwrite it, and releases it after the call. When the call threw, the release was skipped: a try around the call caught the exception, and the global's block kept a count nobody gave back.

• Under rc this applies to any held value.
• Under rc and own it applies to a Shared<T> handle, where Shared.count showed the extra count.

Fix

This only applies when a try in the same function takes the call's exceptions (unwindRegionOf):

• The held value also goes into a slot made at the function's entry. The slot is null except while the call runs: it is stored before the call and nulled after it.
• The handler that runs first starts with a ReleaseSlot of that slot, then nulls it. That handler is chosen in the same order the landing pad uses (TryOpLowering):
  • for a call in the try body: the cleanup, else the catch clause (after the catch variable is set), else the finally;
  • for a call in a catch clause: the try's finally.
• Normal paths through a finally see a null slot, so the release does nothing there.

An exception that leaves the function with no try around the call still leaks the count, the same as owned locals do today.

Tests

• own/own_shared_global_argument_throw.ts runs in the shared_models foreach (gc/none/rc/own) and in the rc/own --verify-ownership list. It covers try/catch, try/finally (caught outside), a block with a local in a try body (a cleanup-only try), a call in a catch clause with a finally, a try in a loop, and the normal path. On main it fails under rc and own.
• Teeth: turning the unwind slot off fails test-compile-rc-… and test-compile-own-own_shared_global_argument_throw (and their jit twins). Without the switch they pass.
• Windows ctest -C Release: 4958/4958. WSL Linux full run: 4940/4940.

🤖 Generated with Claude Code