[ASDAlexander77]

Builds on the scope-exit walk from #591.

Problem

switch (i) { case 0: { using c = new D("case"); } } (or case 0: using c = ...; break;) crashed MLIRGen with 0xC0000005. The cases were generated with a copy of the enclosing context that shared its usingVars/ownedVars, while their names lived only in the switch's symbol table scope, so the dispose at the end of the enclosing scope looked c up and didn't find it.

While testing, a second crash turned up: a break inside an if in a case (case 0: if (x) break;) had no jump target in SwitchOpLowering, so lowering it crashed.

Fix

• The switch has its own scope (GenContext::isSwitchScope, its own lists, parentBlockContext). It is exited at the start of the merge block, which every normal way out reaches: a break at the end of a case, the last case falling off its end, and no case matching. A hoisted using from a case that didn't run has a dispose guard that is still off. A hoisted owned local there is still null.
• break vs continue: new DisposeDepth::BreakScope. An unlabelled break stops at the switch scope, while a continue passes through it to the loop body.
• SwitchOpLowering gives an unlabelled BreakOp nested in a case a jump target, the continuation. That jump skips the merge block, whose cleanup the break already did on its way out.
• A case's lone { } block that declares a using is no longer inlined, so it is disposed before a case falling through runs the next one, as in JS.
• A switch whose cases declare a using (or hold a typed catch, on Itanium) is wrapped in a catch-less TryOp, the same way mlirGenBlockWithUnwindCleanup wraps a block. The switch's storage is hoisted in front of that TryOp, so an exception after a case's using is stored runs the switch's cleanup. Without this the verifier flagged the unwind leak inside a using loop body.
• The safe-cast lambda for switch (o.prop) cleared insertIntoParentScope on the shared switch context instead of on its own copy. Every later case then declared its variables in a per-case scope, and they were gone by the merge.
• Ownership verifier: a hoisted owned slot (consumed VariableOp with no initializer) is now treated as acquired at its stores, not at the hoisted declaration, where the slot is null (VariableOpLowering). This also removes a false report for a break that comes before a declaration in a try body.

Verified

• Dispose order matches under gc, rc and none for: a case with using + break, falling through, no case matched, braced cases falling through, break from an if, a switch in a loop whose body has a using (continue / break / default), and a switch on a property.
• --emit=mlir-affine --verify-ownership is quiet under gc/rc/own.
• New test 00switch_case_scope_exit.ts: compile + JIT, and added to TSLANG_CORPUS.
• Full ctest -C Release -j 8 --timeout 300: 100% passed, 4118/4118 (Windows, VS 2026 Release).

Not fixed here (pre-existing)

• Every scope disposes its usings in declaration order, but JS disposes them in reverse. The test avoids two usings in one scope.
• let s = ...; { using a = new D(); return s; }: the verifier reports s as unreleased on the unwind path, with a plain block as well as with a switch.

🤖 Generated with Claude Code