Hojun-Cho 4d4ad36b70 cmd+selfhost+test: relax alloc-slice element-type pin via LHS retype ...

`alloc([], n)` synthesizes ([]u8 | nomem) at expression level — that's
fine, since the slice form only legitimately appears in let-init
position where the LHS carries the real element type. In clet, after
type-checking the rhs, peel any N_TRYPROP/N_TRYUNW wrapper, match the
alloc-slice AST shape with the same-module shadow gate (from #23),
and retype the call's tagged return to ([]T | nomem) where T is the
declared LHS element. Then assignability sees []T vs []T and accepts.

Cgen N_LET shortcut gains a viatryprop arm next to the existing
viatryunw — on rt_alloc returning null, emits the tagged-return
nomem propagation (MOVQ $nidx, AX; epilogue) instead of exit(1).
nidx comes from cg_tag_for_variant on the enclosing fn's return type,
matching the existing TRYPROP propret path.

Wwstage mirrors all four hunks (check.ww + cgenstmt.ww). Promotes the
previously-silent conf=false skip into a confident accept.

Unblocks #6 (dupall) and lays the path for #4/#7. Byte-identity
holds modulo the pre-existing #44 alloc/rt_alloc symbol divergence.

2026-05-20 01:09:16 +09:00

6.8 KiB

Raw Blame History

View Raw