ΛLogos

Spec conformance — compiler bugs surfaced by test generation (resolution log)

Source references link to 01c0f52b (main) in the Logos repository.

Writing executable conformance tests is an oracle: a rule that cannot be confirmed by an honest runnable assertion is either untestable at language level or the compiler does not uphold it. The rollout flagged 8 candidates; each was investigated with a minimal repro and root-caused in src/compiler. Status:

Fixed (verified green + regression-clean)

Rule id Bug Fix
stmt.let-destruct.move-on-bind Double-free. let (a,b) = tup; dropped the elements twice — lower_let_destruct spilled the source into a temp without marking the named source moved. sema_stmt.cpp: mark the RHS source moved before the spill (mirrors the array-destructure path). Valgrind-clean. Test: tests/spec/pass/stmt_let_destruct_move.logos.
coerce.unsize.return-concrete-to-trait-object Returning Box<Concrete> from -> Box<dyn> segfaulted: the hand-rolled return path didn’t unwrap Box, mis-keyed the vtable, and left the source box un-consumed (double-free). sema_stmt.cpp lower_return: desugar the implicit return to the proven as-unsize cast (consume source + build fat pair + Box drop-glue). Test: tests/spec/pass/coerce_box_dyn.logos.
coerce.deref.box-struct-borrow &Box<dyn Trait> segfaulted — the &-handler had no branch for an owning TraitObject, so &b produced &&dyn (a thin ptr where the callee wanted the fat pair). sema_expr.cpp: add the owning-TraitObject branch (read the fat-pair value, re-type non-owning), mirroring the existing owning-DstRef branch. Test: same file as above.
(bonus) float↔pointer cast f64 as *const f64 (invalid per Rust E0606) was accepted by sema and silently miscompiled (mlir_gen nullptr fallthrough → exit 0, rest of fn elided). sema_expr.cpp: reject float↔pointer casts in the cast-validation block.
expr.method.receiver-multiref-autoderef A method call / deref through an immutable &&T local binding (let r2 = &r1; r2.m()) read one indirection too deep (wrong field / segfault): an immutable &Struct local aliases its pointee address with no own slot, so let r2 = &r1 stored a one-short &P into the &&P slot. (The call-arg f(&r) + &&pat form was already correct — it uses an internally-consistent one-short convention.) mlir_gen_stmt.cpp: a store-side case for let r2 = &<aliased-immutable-ref-local> of &&-type materialises the missing mid slot (r2 → mid → P), restoring the second level. Narrow — doesn’t touch the shared EAddrOf, call-arg, or read paths. Test: tests/spec/pass/expr_multiref_deref.logos.

Not bugs (investigated, no compiler change)

Rule id Verdict
type.typeof.expr Sound. typeof lowering allocates no slot and never poisons sibling typing at any scale; the original symptom was a test-construction artifact (a shadowed binding).
generic.enum-lit.hint-ref-ptr-preference Already fixed on HEAD (niche/tagged repr commits, Jun). Could not reproduce; the rule is sound.
coerce.cast.float-to-float The float↔float width rule is sound at any scale. The real defect was the separate float↔pointer gap (fixed above).
trait.def.vtable-layout-supertrait-closure The supertrait-closure vtable mechanism is correct (verified with a non-colliding trait name; test tests/spec/pass/trait_supertrait_dyn.logos). The reported failure was a trait-name collision: the repro named its trait Sub, which shadows the prelude arithmetic operator trait Sub, so &dyn Sub bound to the wrong trait. That is the pre-existing dyn-local-trait-shadowing gap (adversarial sweep ADV1-H), a separate broad canonicalization fix — not this rule.

Also fixed

Issue Bug Fix
dyn-local-trait-shadowing (ADV1-H) A user trait Sub whose bare name collides with a prelude/imported trait (the ops::Sub operator) registers only under its package-qualified key (B-mv-02), but the dyn path resolved the trait by its BARE name, so dispatch bound to the prelude trait → “trait ‘Sub’ has no method”. Also affected a shadowed supertrait (Sub: Add, both shadowing operators). The in-tree comment predicted a 12-site sweep; the actual fix is 3 scope-aware lookups — the impl/vtable ecosystem is already bare-keyed + target-disambiguated, so only the name-only lookups mis-bound. Swap bare traits_.find → find_trait_iter_scoped (current-package first) in lower_trait_def (sema_decl.cpp), try_method_on_dyn (sema_expr.cpp), and the supertrait-closure walk in trait_vtable_layout (sema_collect.cpp). No-op for non-colliding names. Regression: tests/logos/pass/dyn_trait_shadowing.logos.

Generated during the spec-extract conformance rollout (tools/spec-extract, spec-test.workflow.js). Fixes verified individually + full ctest regression.