Exception unwind releases one interface too few on every cross target

• Track A (the four cross backends' unwind path / exception_emit.inc's cross arms).
• Found 2026-08-21 by the 53-test dyn-array + interface cross differential.

Measured

test/test_interface_arc_exc.pas:

i386, arm32 and aarch64 agree with each other and are short by exactly one, so this is one missing release on a shared path, not three separate defects. riscv32 is short on the reassign line too — that half belongs to [[bug-a-riscv32-drops-interface-releases-in-six-shapes]]; only its unwind shortfall is this ticket.

Why it stays invisible

An unwind leak prints nothing. It surfaces only because this test counts destructor calls, and the test is native-only in the Makefile — nothing runs it for a cross target. The same reason [[bug-a-no-dyn-array-scope-exit-release-on-four-backends]] survived.

Where to start

x86-64's unwind releases the interface; compare its landing-pad cleanup against each cross backend's. The scope-exit arm and the UNWIND arm are separate code paths on every target — the scope-exit one was audited on 2026-08-21 and the unwind one was not, which fits a count that is short by exactly one frame's worth.

Gate

test_interface_arc_exc printing unwind freed=3 under tools/run_target.sh on i386 / arm32 / aarch64; the 53-test cross differential no worse than baseline; self-host fixedpoint + tools/gate.sh quick.

Resolution (2026-08-21)

test_interface_arc_exc now prints unwind freed=3 on i386, arm32 and aarch64, matching x86-64 and FPC. riscv32 is unchanged at 1 and is not this ticket's half (below).

It was not a missing release — it was a missing LANDING PAD

The ticket guessed "compare x86-64's landing-pad cleanup against each cross backend's". There was nothing to compare: the cross backends had no proc landing pad at all. One clause said so, in pasparser_proc.inc:

needsProcCleanupFrame := (TargetArch = TARGET_X86_64) and ExceptionUsed and
                         (not isAsmFunc) and ProcHasManagedLocalCleanup(procIdx, -1);

A proc that holds managed locals and is unwound past — raised below, caught above — never runs its own epilogue and never runs the handler's frame either, so it needs a setjmp frame of its own: chain onto BSS_EXC_TOP at entry, and when a raise longjmps into it, unchain, release, re-raise. x86-64 has had that since exceptions landed. Four targets did not, so every managed local held across a raise leaked there — interfaces, ansistrings, variants, records-with-managed-fields, not just the one the test counts. It printed nothing, which is why it survived: the test that counts destructor calls is wired into test-core only.

Nothing new was written

Every target already emits this exact sequence for a try block (its IR_EXC_ENTER arm) and already has IREmitLeaveExceptionFrames*. A try frame and a proc cleanup frame are the same frame with different landing pads. So the fix is those two pieces called from the proc driver instead of from the statement lowering — lifted into ir_codegen.inc as EmitProcCleanupFrameEnterForTarget / EmitProcCleanupLandingPadForTarget, with the branch-patch encodings (rel32 / bne / cbnz / beq-skip-jal) behind two small helpers so the driver stays target-blind.

The prerequisite was a separate commit: the scope-exit release loop lived inline in each of EmitProcEpilog's six branches, so there was no single thing for the landing pad to call. Extracting EmitManagedLocalCleanupForTarget first (byte-identical, 321153f99) is what made this commit small.

Not covered

• riscv32 now gets the landing pad, but its release arm still has no COM interface case, so it stays at unwind freed=1. That is [[bug-a-riscv32-drops-interface-releases-in-six-shapes]] — the landing pad it now has will start working the moment that arm lands.
• xtensa has no proc cleanup frame. It is the one target whose exception runtime exists only under the Call0 ABI, and its managed-local arm handles AnsiString alone. ESP-campaign work (Track S), stated in the code rather than skipped silently.

Gate

tools/gate.sh quick GREEN. test_interface_arc_exc = native on i386 / arm32 / aarch64. Eight exception tests (test_cross_exception, test_exceptions, test_exception_finally, test_exception_typed, test_managed_exception_cleanup, test_exc_resident_param, test_exceptaddr_b340, test_except_derived_caught_by_base) byte-equal to the native oracle on all three targets. Cross-target breadth is Track T's, against this sha.

Log

• 2026-08-21 — resolved, commit 3af365f10.