own: a const declared from a merge of owning locals borrows them; a value moved into a field of an element read borrows that field (TSLANG 58, 59) - #851

TSLANG_ISSUES 58 and 59, found by compiling powder_sandbox's -mm=own build (native_ts_graphics as its submodule) with the compiler from main. Write-ups are in TSLANG_ISSUES.md.

58: a const declared from a merge of owning locals borrows them

const selected = x2py2gez2pw2 ? (x2gey2 ? tensor0 : tensor1) : (z2gew2 ? tensor2 : tensor3) (ntg core/directx_math.ts) was refused: "merged from several branches; -mm=own cannot prove which one owns it yet". So was a plain const s = c ? t0 : t1.

• Why: each local keeps its value and is released at the end of its scope, so neither value can move into the merge.
• The fix: a let's retain of a merge whose ways in are each a read of an owning local, or null or undefined (slotLoadsOfMerge), is made a receiver of each of those locals. decideSlot then takes the borrow path for them, so rc's retain and release for the const go, and each local's releases and assignments end the borrow.
• Still refused: one of the locals assigned or moved while the const is still used; the const stored, returned or assigned.
• A loop bug, fixed too: takerLoops took a merge's own declaration (its block's first op) as reached again around a loop before the merge is remade. That refused such a merge in a loop's body, and TSLANG 54's const p = c ? mk() : null in a loop as well.
• Result: directx_math.ts compiles under own. With its facts object built, so does core/simple_camera.ts, which had been refused only for calling into a module with no facts.

59: a value moved into a field of an element or field read, then used, borrows that field

const queuedFrame = this.queuedFrames[i]; ... queuedFrame.aabbAttributeBuffer = buffer; const view = buffer ? buffer.createStructuredView(0) : null; (ntg examples/nri/procedural_geometry_nri.ts) was refused: "used here after its value was moved".

• Why: TSLANG 50 and 52 read the field back right after the move only through an object read out of a local or parameter nothing assigns. A const given an element is no local by then; it is the element read itself.
• The fix: right after the store or push nothing runs, so an object read out of an element or a field is the same object there (objectReadableRightAfter). The checks then follow the borrow through that chain.
• Still refused: the element replaced or popped, the field reset, or the holder's field replaced before a use.
• Not loops: under 58 I first took this gap to be about loops; it is the element-read object.
• One site of this shape is left: procedural_geometry_nri.ts:425, which reduced copies of it do not reproduce.

Effect on the sweep (276 sources, every facts object built)

• Sources failing: 141 → 126.
• ntg core/ modules that don't compile under own: 6 → 4. Left: donut_first_person_camera, nri_donut_textures, ray_tracing_extended_scene (needs @keeps/Shared<T> in ntg), rtxpt_game.
• "used here after its value was moved" sites: 49 → 37.
• One unexplained crash: in one sweep run tslang.exe crashed (0xC0000005) on core/ray_tracing_extended_scene.ts. It did not reproduce in about 45 more runs: alone, six at a time, and a full sweep rerun. Each of those gave its usual 5 errors.

Tests

• New tests: 00tslang_issue_58.ts and 00tslang_issue_59.ts, under every model, JIT and compiled.
• New own error tests: own_err_merged_locals_overwritten, own_err_merged_locals_moved and own_err_element_field_moved_replaced.
• ctest (Windows Release): 7257/7257.
• DefaultLib main rebuilt with this compiler, release and debug, all four models. tests.ps1 for each model is running; results to follow in a comment.

🤖 Generated with Claude Code