wcc+w6c_ww: append() spread place-chain sources via cgplaceaddr (#35)

#35 (fold-5 blocker, PG6): the append() spread arm dispatched on
SOURCE node kind — ident-local only; both fold-5 dup shapes
loud-rejected on BOTH stages: deref-spine `append(dup,
(*tsp)[0].caps...)` (add_thread regex.ha:569/572) and indexed
`append(res, ts[0].caps...)` (search ha:820).

Route every non-ident-local source through cgplaceaddr (the
established place choke-point: C1 stores, C1.25 aggregate fields, FA1
append targets, #49 element sources): the source header ADDRESS
resolves ONCE, pre-grow, into a fresh per-site @appsprscr spill, and
every loop iteration re-reads .ptr/.len THROUGH the spilled header
after the grow. This adapts the #49 split ruling to the spread's
full-range copy: the chain's rvalues (deref-root pointer expr, index
exprs) evaluate exactly once PRE-grow — an index reading the dst's
len sees the pre-append value — while the source data base and len
re-derive live each iteration, so a source header aliasing the dst
header re-roots post-realloc. A header reached through a buffer the
grow reallocs reads the STALE copy — bit-identical under the
non-reclaiming rt/malloc, the same documented #49 stale-base hole
(pinned by the spread_selfalias_chain row). Ident-local sources keep
the legacy BP-disp emission byte-identical.

New loudness with the same fix: the spread source's stamped type must
chase to TY_SLICE/TY_STR — a [N]T array ident source previously read
its first 16 DATA bytes as a {ptr,len} header, silently. Array wiring
plus the remaining rvalue sources (CALL, slicing exprs) stay loud,
filed task #27. Global-ident sources now resolve on cstage but are
blocked by a pre-existing wwstage checker reject ("let: not
assignable", task #29) — no dual-stage row until that closes.

806_append_place: reject_spread_src GRADUATES to a runtime row
(spread_place_deref, the old reject source verbatim + readback); new
rows spread_dup_copy (PG6 verbatim: both fold-5 shapes, 56B capture
elements, copy-semantics mutation pin, want 139), spread_place_kinds
(str 24B headers / narrow i32 / empty deref source),
spread_growth_place (40-elem spread crossing cap doublings),
spread_selfalias_chain (source header inside the dst's grown buffer),
and the two new loud-tail rejects (CALL rvalue, [N]T array) pinning
the "#35:" diagnostic on both stages. 87 fixtures green (was 70),
per-row cs/ww asm byte-cmp included.

Unblocks regex fold-5a: add_thread regex.ha:569/572 + search ha:820
spreads go from loud-bound to real spread.
This commit is contained in:
2026-06-04 23:40:31 +09:00
parent 1bcf2726cf
commit 0ce98cc5dc
5 changed files with 449 additions and 81 deletions

View File

@@ -7481,19 +7481,56 @@ cgexpr(Cg *c, Node *n, Local *locals)
* local ident used to fall PAST the spread arm
* into the single-value stores with the
* N_SPREAD node (cstage garbage store; wwstage
* silently SKIPPED it — divergent). Deferred
* source shape, task #37. */
* silently SKIPPED it — divergent). */
if (vn->kind == N_SPREAD) {
if (!vn->lhs || vn->lhs->kind != N_IDENT)
fatal("#34: append() spread source "
/* #35: only a {ptr,len,cap}-headered
* source reads as a header below; a
* [N]T array place IS its storage —
* the ident path used to read its
* first 16 data bytes as ptr/len,
* silently. Loud until wired (task
* #27); str shares the slice header
* layout. */
Type *itu = vn->lhs
? type_chase_named(vn->lhs->type) : NULL;
if (itu == NULL || (itu->kind != TY_SLICE
&& itu->kind != TY_STR))
fatal("#35: append() spread source "
"shape unsupported (rule-7)");
if (localfind(locals, vn->lhs->str) == 0)
fatal("#34: append() spread source "
"ident is not a local (rule-7)");
}
if (vn->kind == N_SPREAD &&
vn->lhs && vn->lhs->kind == N_IDENT) {
int it_off = localfind(locals, vn->lhs->str);
int it_off = 0;
int it_scr = 0;
if (vn->lhs->kind == N_IDENT)
it_off = localfind(locals, vn->lhs->str);
if (it_off == 0) {
/* #35: place-chain source (deref
* spine, indexed chain, global
* ident — the regex.ha:569/820
* dup shapes) resolves its header
* ADDRESS through cgplaceaddr
* ONCE, pre-grow: the chain's
* rvalues run exactly once (the
* #49 split's pre-grow half) and
* every iteration re-reads
* .ptr/.len THROUGH the spilled
* header post-grow (the live
* re-derivation half). A header
* reached through a buffer the
* grow reallocs keeps Hare's
* stale-base hole — see the #49
* comment below. Rvalue sources
* (CALL, slicing exprs) have no
* place — loud, task #27. Fresh
* spill slot per SITE for the
* same nesting reason as
* @apphdrscr above. */
if (!cgplaceaddr(c, vn->lhs, D_BX, locals))
fatal("#35: append() spread source "
"shape unsupported (rule-7)");
it_scr = local_alloc(c, &locals,
"@appsprscr", 8, cg_frame);
ins2(c, A_MOVQ, areg(D_BX),
amem(D_BP, it_scr));
}
int load_op = fldloadop(esub, esz);
/* push counter (i) on stack */
ins2(c, A_SUBQ, aimm(8), areg(D_SP));
@@ -7502,7 +7539,11 @@ cgexpr(Cg *c, Node *n, Local *locals)
char *le = mklabel(c, "spr_e");
label(c, ll);
ins2(c, A_MOVQ, amem(D_SP, 0), areg(D_CX));
ins2(c, A_MOVQ, amem(D_BP, it_off + 8), areg(D_DX));
if (it_scr != 0) {
ins2(c, A_MOVQ, amem(D_BP, it_scr), areg(D_DX));
ins2(c, A_MOVQ, amem(D_DX, 8), areg(D_DX));
} else
ins2(c, A_MOVQ, amem(D_BP, it_off + 8), areg(D_DX));
ins2(c, A_CMPQ, areg(D_DX), areg(D_CX));
ins1(c, A_JGE, abranch(le));
if (el_wide) {
@@ -7522,7 +7563,11 @@ cgexpr(Cg *c, Node *n, Local *locals)
ins2(c, A_MOVQ, aimm(esz), areg(D_AX));
ins2(c, A_IMULQ, areg(D_AX), areg(D_CX));
}
ins2(c, A_MOVQ, amem(D_BP, it_off), areg(D_BX));
if (it_scr != 0) {
ins2(c, A_MOVQ, amem(D_BP, it_scr), areg(D_BX));
ins2(c, A_MOVQ, amem(D_BX, 0), areg(D_BX));
} else
ins2(c, A_MOVQ, amem(D_BP, it_off), areg(D_BX));
ins2(c, A_ADDQ, areg(D_CX), areg(D_BX));
cg_append_slot(c, sn_direct, sn_off,
sn_scr, esz, D_DX);
@@ -7553,7 +7598,11 @@ cgexpr(Cg *c, Node *n, Local *locals)
continue;
}
/* AX = items.ptr[i] */
ins2(c, A_MOVQ, amem(D_BP, it_off), areg(D_BX));
if (it_scr != 0) {
ins2(c, A_MOVQ, amem(D_BP, it_scr), areg(D_BX));
ins2(c, A_MOVQ, amem(D_BX, 0), areg(D_BX));
} else
ins2(c, A_MOVQ, amem(D_BP, it_off), areg(D_BX));
if (esz > 1) {
ins2(c, A_MOVQ, aimm(esz), areg(D_AX));
ins2(c, A_IMULQ, areg(D_AX), areg(D_CX));