A const record whose field is made for it has storage, as a let does (#485)

[ASDAlexander77]

Fixes #485: a const record literal leaked what was made for its fields under rc and own.

The bug

A record literal with a field known only at run time is built in a temporary slot and read out whole, and the literal takes no count for what its fields are given: its receiver does, as a let does with RetainSlot. A const of it was folded into that read unless a field was read out of another object or an owning local (#454, #460), so for a field made for the literal nothing took a count and nothing gave one back.

Peak memory, AOT (measure.ps1), 1M iterations of a function holding const o = { <field>, n: i } and one other owning local:

field	main rc	main own	this PR rc	this PR own
items: [i, 2] (the issue)	81.2 MB	81.2 MB	4.4 MB	4.4 MB
t: [i, [i]]	81.2 MB	81.2 MB	4.4 MB	compile error, see below
items: new Numbers(n)	93.5 MB	93.5 MB	4.4 MB	4.4 MB
inner: { items: [i, 2], k: i }	81.2 MB	81.2 MB	4.4 MB	compile error, see below
s: "a" + i	4.4 MB	35.2 MB	4.4 MB	4.4 MB
items: mk(i)	4.4 MB	compile error	4.4 MB	4.4 MB
c: new C()	4.4 MB	compile error	4.4 MB	4.4 MB

(let o was flat in every row.)

The fix (MLIRGen, readsLiteralSlot)

A const of such a record now has storage, as a let does, when a field owns a block and is made for the literal: an array or tuple built there, a constant array copied for it, a new array, a string, a call's result, a new instance. A value taken out of another record ({ ...o }, ts.ExtractProperty / ts.DeconstructTuple) stays with that record's count, as before, and a literal string is immortal.

Under own, a record whose field is a tuple or record holding an array ({ t: [i, [i]], n: i }) now gives the compile error a let of it already gave on main ("an object literal that holds a value it owns is moved only into a local ..."), where it used to compile and leak.

Tests

• 00const_record_owned_fields.ts (compile, JIT, rc and none corpora): each field kind above, read after heap churn.
• test-rc-const-record-release (const-record-release.cmake): no run sees a leak, so the IR is read - under rc each of that file's seven records has a ts.ReleaseSlot of its tuple slot. Main fails it (arrayField does not release its record's slot).

Gates

Gate	Result
Windows Release suite (ctest -C Release -j 12 --timeout 300, staged default library)	3788 / 3788 passed
Linux (WSL, on main 86a9ec3, ninja release)	3773 / 3773 passed (1 pre-existing disabled test)
own corpus (--emit=llvm -mm=own --no-default-lib, every file of test/tester/tests)	main 459 / 614, this PR 459 / 614 (--opt: 457 / 457). No file goes from ok to failing; 00types, callWithSpread and conditionalTypes2 change only a generated name or a union's member order in their first error
own memory (measure.ps1, AOT)	own_fresh_string, own_return_new, own_try_local, own_fresh_array, own_literals, own_array_param_push unchanged
Default library rebuilt (release and debug) with this compiler, its suite	gc 159/159, rc 158/159, none 158/159 in each of release/debug x compile/jit - the one failure is weakref_basic, gc-only by design (#420)

Closes #485

🤖 Generated with Claude Code