A return, break or continue through a finally disposes and releases the scopes the finally reads after it runs

[ASDAlexander77]

Follows #598, using the same design for disposal and for break/continue.

Bugs

1. return disposed too early. { using r = new R(); try { return 1; } finally { r.read(); } } disposed r before the finally ran, in every memory model. JS disposes it after the finally. The cause: mlirGenDisposable had no finally guard.
2. break/continue released and disposed too early. A jump out of a try in a loop body, or in a block nested in it, disposed and released the scopes around the try up to its loop before it went through the finally copy. Under -mm=rc the finally then read freed strings.

Fix

• scopeExitStopsAtFinallyGuard: both walks, dispose and release, now stop at finallyGuardScope for FullStack, BreakScope and LoopScope. A break or continue notes (kind, label) in GenContext::finallyEscapingJumps, which the body guard and the catch guard share.
• mlirGenScopeExitAfterFinally replaces mlirGenReleaseOwnedOnReturn. At the end of the finally it adds one ts.If true for the return and one for each escaping jump. Each holds the dispose and release of the scopes from the try outwards, using that jump's own walk. Each is tagged RETURN_ONLY_ATTR_NAME, or JUMP_ONLY_ATTR_NAME = "break:<label>" / "continue:<label>".
• IfOp's RemoveStaticCondition skips tagged ifs. Otherwise the canonicalizer that runs before lowering inlines them and the tag is lost.
• TryOpLowering makes one jump copy per (target, key). An unlabelled break and a break outer can share a target and still each get their own scope exit. The group that belongs to a copy is unwrapped (inlineBlockBefore); all other groups are erased. If the wrapper were left in place it would lower to a constant cond_br, and the ownership verifier would see its other edge as a path that never releases.

Tests

00try_scope_exit_through_finally.ts is registered for compile, JIT and TSLANG_CORPUS. It covers:

• return with a using in a block, in a catch, and through two finallys
• break and continue from a nested block and from the loop body
• break and continue with a using
• a labelled continue outer
• a break through two finallys
• a break out of a switch case and out of a labelled block

Values are read after churn().

• JIT under gc, rc and none: all pass. Before the fix, gc failed on the dispose order and rc read freed strings.
• --emit=mlir-affine --verify-ownership -mm=rc: quiet on every test that contains a finally, including the new one.
• ctest -C Release -j 8 --timeout 300: 100% passed, 0 of 4156 failed.

Not fixed (separate, existing bug)

On Windows, a break out of a catch clause crashes (0x80000003) in both JIT and AOT, under gc and none, with or without a finally:
for (...) { try { h(); } catch { if (i == 1) break; } }. The test leaves this case out.

🤖 Generated with Claude Code