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

wcc_ww/cgen: #63 alias-named struct-lit fill via structlookupchain (let-init + sret-return)

Browse Source 
cglet's N_STRUCTLIT init arm resolved the struct by a bare
structlookup(c, sname). For an alias-NAMED literal
(`type rep2 = rep; let r = rep2{id=6}`) the type ref carries the
alias name "rep2" but only the base `rep` is registered, so the
lookup returned nil and the field-fill never fired. The nil then
split by slot size into two symptoms of one root:
  - <=8B: the small-let scalar default zeroed the slot and DROPPED
    the literal (SILENT wrong — the field read 0), and
  - >8B: no fill arm matched, falling to the cglet "unhandled rhs
    shape" LOUD (task #7/rule-7).

Route the arm through structlookupchain (the #92/W2 SSoT already
adopted at cgenstmt:1974/:2687), which chases the alias chain to the
base struct. trefn (rhs.lhs) is already the N_IDENT/N_TNAME type ref
structlookupchain accepts, so the bare sname extraction is dropped.
cstage operates on the resolved Type* via type_chase_named and was
always correct: ww-only align-UP, cs UNTOUCHED.

ROUTED (the two reachable silent sites, one class):
  :2421  local N_STRUCTLIT let-init — the #63 repro hits it for
         both the <=8B silent-zero and the >24B loud symptoms.
  :1250  >24B sret RETURN twin (reviewer-63). sretretsize chases
         the alias for the size GATE so this sret arm fires, but
         the fill used the same bare structlookup(sname) — for an
         alias-named >24B literal it returned nil and the fill was
         SKIPPED, so the callee returned an uninitialised sret
         buffer (SILENT wrong, runtime-0; cs correct). Same root,
         same symptom, sibling site → folded by construction.
DECLINED (traced, not blind-routed; rule-11 + the #101 precedent):
  :2625  N_IDENT struct-copy — also bare-structlookup but the copy
         falls through to a generic path byte-identical with cstage;
         both stages run correct. The post-copy field-READ diverges
         (cs direct-offset vs ww LEAQ-indirect) = the #81/#65 alias
         field-read class, out of #63 scope.
  :2511  N_CALL struct-recv — blocked UPSTREAM by the aggregate-
         return shape (#272/#277); ww louds at the sender.
  :1363  <=24B register RETURN — alias case louds via the same
         scalar-default catch (#277), not silently wrong.
The 2 already-chasing sites (1974/2687) untouched.

CONVERGENCE: m3_letinit_typed + m3_letinit_untyped (ww silent-zero ->
6/6 byte-id) + m6_letlit_alias (ww loud -> 7 byte-id) + sret_return_-
alias32 (ww silent-0 -> 10 byte-id), plus a non-alias control
(no-regress). Bootstrap byte-id NEUTRAL (selfhost has no
alias-struct-litinit/return; all 4 selfhost tools cs==ww confirmed).

Test: 944_alias_structlit_init_run (5 rows x cs-run + ww-run +
cs==ww byte-id = 15 checks), Makefile-wired.
This commit is contained in:
Hojun-Cho
2026-06-06 10:36:16 +09:00
parent 45f5415209
commit 3546673756
5 changed files with 329 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
Makefile
View File

@@ -303,6 +303,7 @@ TESTS = $(BIN)/test_smoke $(BIN)/test_lex $(BIN)/test_parse $(BIN)/test_check \
$(BIN)/test_subslice_ptresz_run \
$(BIN)/test_deref_slice_store_run \
$(BIN)/test_alias_decl_order_size_run \
$(BIN)/test_alias_structlit_init_run \
$(BIN)/test_alias_accept_run \
$(BIN)/test_alias_idx_family_run \
$(BIN)/test_alias_amp_idx_run \
@@ -1473,6 +1474,12 @@ $(BIN)/test_alias_decl_order_size_run: test/wcc/944_alias_decl_order_size_run.c
$(LIB)/libwwrt.a | $(BIN)
$(CC) $(CFLAGS) -o $@ $<
$(BIN)/test_alias_structlit_init_run: test/wcc/944_alias_structlit_init_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_alias_accept_run: test/wcc/944_alias_accept_run.c \
$(BIN)/ww $(BIN)/w6c $(BIN)/w6a $(BIN)/w6l \
$(BIN)/ww_ww $(BIN)/w6c_ww $(BIN)/w6a_ww $(BIN)/w6l_ww \