← board

The scope-exit release matrix still has four holes

The matrix

Now that all six arms live in one procedure, the gaps read off in one grep:

arm x86-64 i386 arm32 aarch64 xtensa riscv32
COM interface yes yes yes yes — yes
static array of managed yes yes yes yes — yes
ansistring yes yes yes yes yes yes
variant yes NO yes yes — yes
promotable int yes NO NO yes — yes
record w/ managed fields yes NO yes yes — yes
dynamic array yes yes yes yes — yes

Four holes: i386 × {variant, promotable int, record}, arm32 × {promotable int}.

Why it matters and why nothing catches it

A missing arm is a leak per call, and a leak prints nothing — which is how the dyn-array arm stayed missing on four backends, how the static-array arm stayed missing on three, and how riscv32's interface arm stayed missing on one. The tests that catch these are the ones that COUNT destructor calls, and they are wired into test-core (native) rather than the cross suites.

So the fix needs its own evidence: either a counting test built for a cross target, or -dPXX_HEAP_DEBUG before/after.

Where to start

Copy the arms from aarch64's block — it is the only cross target with all six — translating the register moves. Each is 6-10 lines: address of the slot into the first argument register, a descriptor into the second where the helper takes one, EmitCallProc. The helpers (PXXVarClear, PXXPromoClear, PXXRecordRelease) are ordinary Pascal and already exist.

xtensa's nearly-empty row is deliberately NOT in scope: its exception runtime exists only under the Call0 ABI, so it is Track S.

Gate

A counting test showing the freed count matching native under tools/run_target.sh for i386 and arm32; self-host fixedpoint + tools/gate.sh quick.

Resolution (2026-08-21)

All four holes closed. The matrix now reads yes across every hosted target:

arm x86-64 i386 arm32 aarch64 xtensa riscv32
COM interface yes yes yes yes — yes
static array of managed yes yes yes yes — yes
ansistring yes yes yes yes yes yes
variant yes yes yes yes — yes
promotable int yes yes yes yes — yes
record w/ managed fields yes yes yes yes — yes
dynamic array yes yes yes yes — yes

Each added arm is aarch64's, translated: slot address into the first argument register, descriptor into the second where the helper takes one, EmitCallProc. The helpers (PXXVarClear, PXXPromoClear, PXXRecordRelease) already existed.

The evidence problem, and what solved it

The ticket asked for "a counting test built for a cross target, or -dPXX_HEAP_DEBUG before/after". Neither is available for these three arms: counting needs a destructor, and a Variant, a record and a promotable int have none. That is precisely why these holes lasted — every test that ever caught one of these counted destructor calls, so only the class-shaped arms were ever watched.

test/test_managed_local_release_reuse.pas counts nothing and observes the allocator instead. pxx's free list hands a released block straight back for the next same-sized request, so a routine that allocates one managed local per call returns the SAME payload address every call — unless the local leaks, in which case each call takes a fresh block and the address marches upward. First call versus fortieth is the whole assertion: no destructor, no counter, no heap instrumentation, and it works for every managed kind.

It reproduced the bug before the fix, which is the part that matters:

i386   before:  rec first=FALSE   var same=FALSE      (arm32/aarch64: TRUE)
i386   after:   ok 5 / 5          (native, arm32, aarch64: ok 5 / 5)

One trap worth recording: the first draft had the static-array round allocate all THREE elements, and it failed on x86-64 with nothing leaked. The free list is LIFO per size bin, so freeing three blocks and requesting three hands them back reversed — the address differs every call. The probe is only valid with ONE allocation per round.

Wired into the core list and the i386 / arm32 / aarch64 target blocks as an output comparison against the x86-64 build, which is the right oracle here since the arms differ per backend.

Not covered

Gate

tools/gate.sh quick GREEN (self-host fixedpoint 103s). test_managed_local_release_reuse = 5/5 on native, i386, arm32 and aarch64, and FALSE→TRUE on i386 for the two arms that were missing. Cross-target breadth is Track T's, against this sha.

Log