Hojun-Cho/ww
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

wcc: accept bare &fn into a fn-pointer-alias slot via a caller-site gate (#206)

Browse Source 
A bare `&fn_name` was not assignable into a `*reader` / `(*reader | void)`
vtable field without an explicit cast: cstage type_eq on TY_NAMED is
pointer-identity, so a structural `*fn(...)` referent never matched the named
`*reader` variant; wwstage accepted it only via an accidental catch-all
leniency. harec accepts bare &fn through hint-directed alias adoption at the
address-of site (check.c:3594-3626) while keeping pointer assignability
strictly nominal (types.c:1039-1066), so a materialized `*fn` value never
launders across alias names.

Mirror that decision without threading a type hint through the bottom-up
cexpr: keep type_assignable / isassignable fully nominal, and add a
caller-site helper (assignable_addrfn) at the assignment boundaries
(let-init, struct-literal field-init, assign, return, call-arg, array
element) that accepts iff the rhs is a DIRECT &-of-fn-ident and the
destination (or exactly one tagged variant) is a pointer-to-fn-alias whose
underlying fn signature structurally matches. A materialized `*fn` value, a
distinct same-signature alias, and an ambiguous multi-variant target all stay
rejected. Both stages share the rule; wwstage's lenient pointer-fn punt
becomes a confident reject. ww has no methods, so a `value.leaf` slot is only
ever a fn-pointer field and this never over-admits.

The tightening surfaced a wwstage typeeqast gap: a TY_FN result that is a
tuple (`*fn(...)(i32,i32)`) compared false where cstage type_eq handled it,
newly rejecting a legitimate structural assign. Add the N_TTUPLE structural
case (rule-10), restoring test 766.

cgen-neutral (the cast was a no-op reinterpret); pre/post bootstrap .s
zero-delta. Test 783 covers the positive paths (incl. a byte-id-clean
three-field-vtable dispatcher) and the negatives. Tagged-slot negatives
(ambiguous / tagged-laundering) are rejected on cstage but wwstage's separate
`(X|void)` void-variant leniency (#214) still admits them; 783 pins them
cstage-only, to graduate when #214 closes (required before wwstage becomes
the authoritative selfhost checker).

Note: `make clean && make test` is RED at HEAD on 4 alloc fixtures
(700/748/758/915) via a pre-existing clean-build defect (#215, malloc vs
rt_malloc); identical with or without this change, so bisect-clean for #206.
This commit is contained in:
Hojun-Cho
2026-05-29 18:05:53 +09:00
parent 4e1181fd8c
commit f6ac7fb2f8
6 changed files with 908 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
Makefile
View File

@@ -331,6 +331,7 @@ TESTS = $(BIN)/test_smoke $(BIN)/test_lex $(BIN)/test_parse $(BIN)/test_check \
$(BIN)/test_fmt_vstream_mods_run \
$(BIN)/test_fmt_vstream_compositions_run \
$(BIN)/test_fieldfn_leaf_collide_run \
$(BIN)/test_amp_fn_assign_run \
$(BIN)/test_bufio_vstream_run \
$(BIN)/test_log_vstream_run \
$(BIN)/test_use_promote_alias \
@@ -716,6 +717,14 @@ $(BIN)/test_fieldfn_leaf_collide_run: test/wcc/782_fieldfn_leaf_collide_run.c \
$(LIB)/libwwrt.a | $(BIN)
$(CC) $(CFLAGS) -o $@ $<
# #206: bare &fn assignable into *<fn-alias> / (*<fn-alias>|void). Both
# stages (cstage + wwstage byte-id); single-file probes, no lib imports.
$(BIN)/test_amp_fn_assign_run: test/wcc/783_amp_fn_assign_run.c \
$(BIN)/ww $(BIN)/w6c $(BIN)/w6a $(BIN)/w6l \
$(BIN)/ww_ww $(BIN)/w6c_ww $(BIN)/w6a_ww $(BIN)/w6l_ww \
$(LIB)/libwwrt.a | $(BIN)
$(CC) $(CFLAGS) -o $@ $<
$(BIN)/test_fmt_vstream_mods_run: test/wcc/780_fmt_vstream_mods_run.c \
$(BIN)/ww $(BIN)/w6c $(BIN)/w6a $(BIN)/w6l \
$(BIN)/ww_ww $(BIN)/w6c_ww $(BIN)/w6a_ww $(BIN)/w6l_ww \