[ASDAlexander77]

Phase 6 of -mm=own: the corpus report, and the shapes it says matter. The plan is docs/superpowers/plans/2026-10-02-own-phase-6.md; the results are in spec §20.

Rebased on main after #443 and #444 merged: 5 commits. On the rebased branch, the Release suite passes 3415 of 3416 (the same stale default-library test), and the own corpus is unchanged at 402 compiling files. The corpus now has 592 files; #443's new 00iterator_methods is rejected under own, as #443's comment documents.

The report

After 7b, 227 of the 591 corpus files fail under own. 137 of those stop at the catch-all "takes a second reference", which names no rule. A temporary dump at the catch-all's four report sites (kept out of the tree) split the 137 into shapes. Spec §20.1 has the tables, including how many files have each shape as their only blocker.

What this PR accepts

• A global's initializer. It runs in the global's region, which no function's inference sees, so rc's retain for its ts.GlobalResult reached the lowering's guard ("left a retain behind", 8 files). The signature pass now erases that retain when three things hold:

  • the value is the only reference to its block;
  • it is fresh or constant data;
  • every use ends in the ts.GlobalResult: directly, through a view, as an array-literal element, or copied into an object literal's box.

  Any other retain there is reported by the global's name.

• Data nothing owns (isConstantData, OwnershipFacts.h): a number or null in a union, a ts.Constant or a field of one, a function named by a ts.SymbolRef, and a tuple of these.

  • Its retain is erased.
  • A ts.RetainSlot only qualifies while nothing is stored into the slot or its fields.
  • It also has no use after a move: under --opt, CSE merges equal literals.

• A record literal made for an owning local, let a = { item: new C() }, and a = {...}. This is 7a's "record built to be returned", with an owning local in place of the result slot.

For review

• The two ownsHeaps disagree by design. isConstantData lives in OwnershipFacts.h and both passes read it. The header's ownsHeap looks at the type only; the inference pass overrides it for a ts.New seen as its object. That disagreement produced a regression while this PR was being built: 00object, one object literal stored into two records, which 7a rightly rejects, compiled again. Hence:
  • an explicit ts.New exclusion, pinned by the new negative own_err_literal_two_records;
  • an exclusion for values a ts.NewInterface view boxes. For own_err_interface_two_views, each exclusion alone keeps the test rejected.
• The corpus script compiles without --opt, but the test runner uses --opt --opt_level=3. The gate (test-runner AOT and JIT on every newly compiling file) caught 01optional, which only failed under --opt. The final count is the same either way.

Verification

• Release suite: 3409 of 3410 pass. The one failure, test-compile-gc-defaultlib-collector, is the stale installed default library, as on -mm=own phase 5a: a closure that does not escape borrows what it captures #436-A global declared in top-level code owns what it is given (rc; -mm=own phase 6, PR 1) #444.

• WSL/GCC: builds, and the 2066 tests run there pass (own, rc corpus, none corpus).

• Own corpus: 364 → 402 of 591 including A global declared in top-level code owns what it is given (rc; -mm=own phase 6, PR 1) #444 (402 with --opt too), none lost. Re-gated on the final build, all 38 pass test-runner -mm=own, AOT and JIT, with gc's output.

• First errors: the catch-all falls 137 → 101, and "left a retain behind" 8 → 0.

• Memory: a loop building and dropping let a = { item: new Leaf(i) }, AOT, measure.ps1, in MB. This is the only rule here that erases a release.

  iterations	gc	rc	none	own
  20000	5.9	4.2	7.7	4.2
  200000	5.9	4.2	35.2	4.2

• Teeth: each rule was undone with a temporary switch, and its test fails. That covers the global erase, the constant-data erase, the constant-data no-use-after-move check, the ts.New exclusion and the record rule. The boxing exclusion is a second guard (above).

Tests

• Positive: own_global_init, own_constant_data, own_record_local.
• Negative: own_err_literal_two_records.

Not in this PR (spec §20.4-§20.6)

• Next PR: named messages for the shapes the dump recognises (slot reads, conditional results, nested records), and a check of the use-after-move and borrows-a-field buckets for false positives.
• Known limits:
  • a borrow of data nothing owns, past its owner's overwrite;
  • nested record literals;
  • arrays of record literals;
  • a conditional's result.
• Found, not fixed (every model):
  • A global initialized from another global reads garbage (let b = a at module level). With a class, the program faults.
  • A module with a top-level statement, imported as source, fails with invalid LLVM IR.
  • With -shared, an exported const from such a module can't be resolved.

🤖 Generated with Claude Code