cstage+test: SK_USE→SK_X promotion sets use_alias

In cmd/wcc/check.c the pass-1.5 SK_USE→SK_DEF/SK_FN/SK_VAR promotion
sites forgot to set prev->use_alias = 1 when the imported module's
top-level decl shadowed the SK_USE leaf in flat scope. Downstream
dot-prefixed lookups (resolve_typename L77, N_DOT L709) gate the
module-head walk on (SK_USE || use_alias), so `mod.flag` resolution
fell through to "unknown type". The SK_TYPE precedent at L1660 had
the line; the three sister sites at L1709/L1722/L1736 now do too,
in the same one-line shape and field-set order.

The wwstage selfhost/cmd/wcc/check.ww uses coexistence rather than
in-place promotion: SK_USE and same-leaf SK_TYPE/FN/DEF/VAR live as
separate entries differentiated by sym.mod, and scopelookupinmodule's
mod-filter already disambiguates dotted lookups — no use_alias flag
needed, so the cstage bug is structurally non-reachable there. An
architectural note at installdecl documents this divergence-by-design
and warns against porting the flag (adding a field to `sym` changes
its size and risks the wwstage cgen amalloc-undersize trap).

Audit covered every SK_USE→SK_X promotion path in check.c (4 sites:
SK_TYPE already-correct as precedent, SK_DEF/SK_FN/SK_VAR fixed). The
surfacing case was lib/fnmatch: `fn fnmatch(...)` shadows the SK_USE
leaf, so `fnmatch.flag` failed in worker-fnmatch's WIP — that test
(972_fnmatch_run) now flips PASS as live integration proof.
test/wcc/699_use_promote_alias.c pins all four rows with a single
table-driven driver (type/fn/def/var → use mod; let m: mod.flag =
mod.flag.A; return m: i32, expecting exit 42 per row). 995_self_rebuild
byte-identity holds.
This commit is contained in:
2026-05-15 15:06:37 +09:00
parent 86de58bc6c
commit 7f60ebbe44
11 changed files with 261 additions and 0 deletions

View File

@@ -94,6 +94,20 @@ fn declmod(file: *node, d: *node) str = {
// installdecl — install the top-level decl's name into the top scope.
// We don't compute its type yet (that's the resolve pass) — just bind
// the name so forward references resolve.
//
// Architectural note: wwstage uses COEXISTENCE rather than the cstage
// promote-SK_USE-in-place approach in cmd/wcc/check.c. SK_USE and any
// same-leaf SK_TYPE/SK_FN/SK_DEF/SK_VAR live as separate entries in
// the same scope-bucket, distinguished by `sym.mod`. The dot-prefix
// lookup in resolvewalk + scopelookupinmodule's mod-filter already
// disambiguate `fnmatch.flag` against an `fn fnmatch(...)` of the same
// leaf — no `use_alias` flag needed. So the cstage L1722-class bug
// (promotion missing use_alias) is structurally non-reachable here.
// Don't port the use_alias flag from cstage without first re-reading
// the architecture: adding a field to `sym` changes its size and risks
// the wwstage cgen amalloc-undersize trap (rob-pike). #11 (wwstage
// checkfile pass) will reconsider this when wwstage grows a real check
// pass on the cgen path.
fn installdecl(c: *checker, file: *node, d: *node) void = {
if (d == nil) { return; };
let k: nkind = d.kind;