A break, continue or return out of a catch clause ends the catch on every path out of it

[ASDAlexander77]

A break or continue out of a catch clause crashed with 0x80000003 on Windows (JIT and AOT, gc/rc/none, with or without a finally), and so did a return out of a catch that could also fall off its end:

for (let i = 0; i < 3; i++) {
    try { h(); }
    catch { if (i == 1) { break; } }
    trail += "a" + i + ";";
}

There were two defects.

MLIR (TryOpLowering). A break/continue out of a catch was marked as leaving the catch only when the try had a finally. Without one, the jump ended no catch. A jump out of a catch nested in another catch ended only one of the two. TSContext::catchesLeft now counts how many catch clauses a return, break or continue leaves, and the lowering emits one end-of-catch for each. The return/jump in #599's catch-all finally copy uses the same counter.

Win32ExceptionPass. Each catch region got exactly one end. The scan follows block layout and closed the region at the first __cxa_end_catch. Any other way out was left as a bare branch out of the catchpad, and the calls on that path had no funclet bundle, so WinEHPrepare turned them into unreachable code. The new findOtherCatchExits walks each catch clause by control flow from its pad. It walks nested catches first and continues after their ends. It collects every end marker and the calls on the way to each one, and each extra end gets its own catchret. The same walk fixes the shape #599 left open: a finally that returns or breaks on only some paths (finally { if (c) return s; ... }). For that shape, the rethrowing run-to-end exit is now always the region's main end. An invoked end marker is also recorded among the region's invokes, so the rethrow goes to the enclosing handler.

New test 00try_jump_from_catch.ts (compile, JIT, TSLANG_CORPUS) covers:

• break, continue and return out of a catch, with and without a finally
• a finally that reads a local of the block around the try
• a typed catch
• a catch nested in another catch
• a loop inside a catch
• a labelled continue
• a finally that returns or breaks on some paths only, with and without an enclosing try

Verified on Windows:

• the new test passes under gc/rc/none, with and without --opt
• --emit=mlir-affine --verify-ownership -mm=rc reports nothing on it
• ctest -C Release -j 8 --timeout 300: 4168/4168 passed

The Itanium path also emits the extra __cxa_end_catch calls, but I have not run it locally; the Linux CI job is its check.

🤖 Generated with Claude Code