A [N]tagged-union array-literal element fell through the is_agg
multi-word-copy path (STRUCT/ARRAY/TUPLE/str/slice only) to the scalar
1-word store: the raw value landed in word 0 (the tag slot) with no tag
written and no payload boxed, so a later match found no variant. Both
stages under-copied identically, so the copy-depth bug was byte-id-blind
— a stride-only fix would still store 1 word and pass the gate green on
both-wrong.
Route each tagged element through cg_widen_tagged_store / the N_LET "BP"
tagged-store wrapper — the same choke-point let-init, vararg gather and
struct-field stores already use — so boxing, tag-remap and zero-pad-to-
slot come for free. esz now comes from the stamped slot size (rule-13);
the wwstage narrow override only covered widths 1/2/4, leaving a 16/24B
tagged element on the wrong 8-byte sentinel stride. rule-7 loud-stops
the unwired `[N]tagged=[x...]` repeat-fill (the widen call consumes the
node and trashes AX).
test/wcc/685: table-driven runtime readback (106/42/13) + a build-fail
row for the repeat-fill loud-stop, both stages.
`t[i].cap` (t a `[N][]u8` / `[N]str`) miscompiled in BOTH stages,
divergently — the read-side sibling of #20's store fix. cgexpr on the
indexed element leaves the full {ptr,len,cap} header (AX/BX/CX via
cgslicehdr), but the `.cap` field-selector never shuffled CX→AX:
cstage's typed pseudo-field else-branch handled only .ptr/.len, so
`.cap` fell through returning AX=.ptr; wwstage's cgdot non-ident
catch-all likewise handled only .ptr/.len, emitting no read (stale AX).
`t[i].len` already worked (BX→AX shuffle) — only `.cap` was missing.
Fix mirrors the .len shuffle: add the .cap CX→AX arm in both stages.
The shuffle fires ONLY for a typed slice/str base (TY_SLICE/TY_STR
after NAMED-chase); an untyped str literal (`"abc".cap`) leaves only
AX=ptr/BX=len and must return AX unshuffled — keeping the wwstage
catch-all byte-identical with cstage, whose cap-shuffle lives in the
typed branch, not the untyped catch-all.
Validated direct `t[i].cap` (slice + str, elements 0/1) against the
whole-element-copy oracle (`let q=t[i]; q.cap`, made correct by #20),
plus .len-after-index regression pins, in test 683; dual-stage runtime
+ byte-id (36/36 ok). combined.ww regenerated.
A `let t: [N][]u8 = [a, b]` / `[N]str` literal init lowered each
element's {ptr,len,cap} header into AX/BX/CX (cgexpr) but stored only
some words: a slice element fell through to the scalar 1-word MOVQ
(dropping .len AND .cap), a str element stored 2 words (dropping .cap,
latent). Each element is 24B (post-#1) and must be copied whole.
wwstage was worse — a slice element matched no esz branch, so esz
stayed the 8 sentinel: the per-element stride collapsed (element i+1
overwrote element i's tail), the -96-vs-80 cs!=ww frame divergence.
This is the str/slice arm of the #270 aggregate-element-store family.
struct/array/tuple already copy correctly via the #270-1c is_agg
multi-word path; str/slice were the documented follow-up (cgen.c:9037,
cgenstmt.ww deferral). They can't join is_agg (that path word-copies
from a source slot and rejects non-ident/structlit elements, whereas
str/slice elements are commonly exprs cgexpr lowers into registers) —
the correct mechanism is the existing register header store, extended.
Fix (BOTH stages, converged byte-identical): cstage adds
is_slice_el = type_isslice(esub) and stores 3 words (incl CX->base+16,
the cap) for `is_str_el || is_slice_el`, in the main loop and the
repeat-fill. wwstage adds isslicel (esubti.kind == TY_SLICE -> esz =
esubti.size, fixing the stride) and the matching 3-word store. Closes
[N][]u8 (the bug) and the latent [N]str cap-drop in one branch.
The latent str cap-drop is now stored, but the indexed-element `.cap`
READ (`t[i].cap`) stays broken — a distinct cgindex/dot-selector bug,
cs!=ww divergent, filed as task #13. The new test validates the stored
cap via a whole-element copy (`let q = t[i]; q.cap`), which reads
through the correct ident-load path. [N]tagged literal init is the
remaining sibling (is_agg excludes TY_TAGGED), task #12.
Test 683_arr_strslice_elem: table-driven, dual-stage runtime + asm
byte-id; slice/str .len, 3-element stride-24, cap-via-copy, .ptr deref,
plus a [N]struct regression pin proving the is_agg path is untouched.
`len(xs[i])` over a [N]str/[]str (and []T slice) element returned the
element's .ptr, not its length, on BOTH stages (shared gap, not rule-10):
the len() builtin had no N_INDEX arm, so it fell to the bare-cgexpr
fallback, where the N_INDEX str/slice load (cgslicehdr) leaves AX=.ptr,
BX=.len, CX=.cap — and len() returned AX (the ptr) as the length.
Add an N_INDEX arm gated on a (TY_SLICE||TY_STR) element in both stages:
cgexpr the element, then MOVQ BX,AX to shuffle the len word into the
result reg — the same shape as the #14 .len pseudo-field fix. Byte-id
neutral (no bootstrap source uses len(indexed-element)); regenerated
w6c + wwdump combined.ww. New 802_lenidx_run pins runtime + cs==ww.
A module-level `let xs: [N]str = ["a","b",...];` static init emitted no
.data: a str element carries a ptr->rodata relocation, not just bytes, so
it fell through the byte-only array-emit path and left the table symbol
undefined (w6l: undefined reference). Shared gap on both stages, not
rule-10.
emit_strarray_data / emitstrarraydata apply the scalar-str-global pattern
per element at offset idx*esz: a DATAW row of {0-ptr placeholder, LE len,
cap} plus a per-element DATAR sym+idx*esz,_S_n reloc. let_pre_intern /
letpreintern pre-intern each element strlit so the _S_ rodata rows precede
the DATAR references. Stride routes through etype->size (rule 13). Scoped
to the DATAW (`let`) directive: A_DATAR requires a DATAW holder, so
`def [N]str` and str-in-aggregate stay a filed follow-up.
919_strarray_static_run pins runtime (len-sum, element .ptr deref, var
index, empty slot, repeat suffix) + cs==ww byte-id. w6c + wwdump
combined.ww regenerated.
A string literal is TY_UNTYPED_STR, not TY_STR, so `"abc".len` missed
the typed slice/str pseudo-field gate in cgen.c's N_DOT and fell to the
final base-eval fallback, which left AX=.ptr — `.len` returned the
pointer instead of the length. wwstage's cgdot catch-all already did the
BX->AX shuffle, so the two stages diverged (rule-10). Align cstage UP:
the N_DOT fallback emits MOVQ BX,AX for `.len`. `.ptr` is unchanged
(already returned AX); `.cap` deliberately not added (wwstage catch-all
is ptr/len only — mirror exactly).
byte-id was blind here: no bootstrap source uses literal `.len` (lengths
are hardcoded around literals), so the gate never exercised it. New test
801 pins both dimensions (cstage run + cs==ww byte-id) over
len/empty/multibyte/ptr-deref/arg-passthrough rows.
Review fixes for the #272 fold (reviewer272b gate; rob+ken ruling). Bundled
because the wwstage catch-all message carries the citation and the combined.ww
regen covers both .ww edits.
- wwstage cgreturn close-by-construction catch-all keyed on the SYNTACTIC
return-type node (N_TARRAY / N_TNAME+structlookup), so a named-alias
aggregate return type (type a=[N]T / type a=struct) bypassed both the
handling arms AND the loud-stop, falling to the scalar default = silent
segfault/truncation; cstage (type_chase_named at all 4 N_RETURN sites)
stayed correct. Re-key the catch-all on the RESOLVED tinfo (chase
TY_NAMED -> TY_ARRAY/TY_STRUCT) so wwstage LOUD-STOPS (rule 7) instead of
miscompiling. cstage stays correct; the full wwstage tinfo-kind dispatch
(align UP, byte-id) is #277. Established wwstage-stricter divergence
(cf #264), no bootstrap consumer (990-997 green).
- #276 citations at-site (both stages): the cstage >24B array-literal return
loud-stop and the <=24B STRUCT global-receive residual now cite #276. The
wwstage >24B array-literal routes through the tinfo-keyed catch-all
(#272/#276/#277). Correction: ALL <=24B struct globals truncate
symmetrically (byte-id-clean), not only float-bearing -- #276 broadened.
- Cosmetic: fix a double-encoded U+2264 (mojibake) in the cgen.c commit-2
comment.
combined.ww regenerated (#110).
The caller-half of the global case: `g = mk()` into a GLOBAL array
stored only the first word — a ≤24B reg-return landed `MOVQ AX, g(SB)`
(8 of 24 bytes); a >24B sret-return hit the #220 sret-to-symbol gate
which was TY_STRUCT-only and fell through to the same truncation.
≤24B: the local aggregate-receive arm was `off != 0`-only, so a global
array fell to the scalar IDENT store. Add a global ARRAY arm — LEAQ
name(SB), DI then store the full+tail words from AX/DX/CX (an array is
never float-class, so AX/DX/CX is always the transport; no `g+8(SB)`
operand form exists). Mirrors the str/slice global arm.
>24B: add TY_ARRAY to the #220 sret-to-symbol gate (cg_sret_dest_sym /
sretdestnode) — the callee writes the whole array through RDI.
A ≤24B STRUCT global receive can be float-class (X0/X1, not AX/DX/CX),
so it is left at its pre-existing symmetric behaviour — no consumer.
949_aggret_source_run gains global_recv (c → 15) and global_recv_sret
(>24B → 22), both with per-row byte-id.
The N_RETURN aggregate arms gated the return source on N_IDENT ||
N_STRUCTLIT; every other aggregate rvalue (array literal, o.field N_DOT,
a[i] N_INDEX, *p deref) fell through to the scalar-AX default = a silent
8-byte truncation. Both stages emitted byte-IDENTICAL wrong asm, so the
byte-id gate could not catch it (#263 class) — the fix converges on the
runtime oracle.
Mirror the arg-side closure #271 landed: both arms (≤24B @retscr and
>24B sret) now funnel N_ARRLIT through the literal element fill and
N_DOT/N_INDEX/deref through aggarg_srcaddr + the #265/#268 whole-
aggregate copy. Type-agnostic, so struct AND array returns are closed.
A close-by-construction loud-stop (rule 7) guards any future unhandled
aggregate source from reaching the scalar default.
Closes the callee-half of (b)/(c) and the addressable siblings. The
g = mk() global-receive caller-half is commit-2.
949_aggret_source_run pins the class: array-literal / N_DOT / N_INDEX /
deref / named-ident control / >24B-sret-deref / struct-field / struct-
deref, each summing all members (full readback) with per-row byte-id.
Passing an aggregate BY VALUE as a call argument worked ONLY for a ≤16B
struct from an IDENT source; every non-ident source — CALL mk(), N_DOT
o.f, N_INDEX a[i], DEREF *p — and every array / >24B-struct (even as an
ident) fell to the scalar default: one PUSHQ for a multi-word aggregate,
stack-imbalancing against the type-based multi-word drain. cs!=ww, both
garbage (f(mk()) cs4/ww236, f(o.f) cs8/ww108, f(a[i]) cs4/ww28, f(*p)
cs4/ww140; arrays + 32B sret struct same).
The arg-pass twin of the #265/#268 let-init copy. A new aggregate-arg
push arm materialises the source into the arg convention: the source
ADDRESS in SI (ident LEAQ / deref operand / dotchainaddr #253 /
&base[i] spine #252-270) then its ceil(sz/8) words pushed high→low; a
CALL receives first — ≤24B in AX/DX/CX pushed straight, >24B sret'd
into a per-fn @aggargscr then pushed from there. The pop-forward drain
gained a matching array / >16B-struct arm and the callee prologue an
is_bigagg receive (ceil(sz/8) GP eightbytes), so caller and callee
agree on the multi-word layout. The ≤16B-struct-IDENT fast path is
untouched (byte-id preserved).
The new-arm exclusion is TYPE-keyed (the stamped tinfo, mirroring
cstage node_isstructarg over args[i]->type), not the name-keyed
structparamsize — a name-keyed gate re-opened the #211/#13 cross-module
same-leaf collision (784 symmetric: an 8B `sa.s` struct whose
name-resolution collides with `sb.s = *vtable` would miss the struct
fast path and wrongly enter the new arm, diverging from cstage's
1-word push). A float-bearing ≤16B struct from a non-ident source
loud-stops in both stages (the #165 SSE eightbyte transport the GP
push/drain can't model; out of scope). A const array/struct `def`
global as an aggregate arg is aligned DOWN to the leaner wwstage
(both loud-stop) per rule-10.
#110: cgen is compiler-imported by w6c + wwdump — main.combined.ww
regen'd for both.
949 rows: arg_{struct16,arr16,struct32}_{call,dot,idx,deref,ident},
full member readback (struct 16B reg-class + 32B sret-class + array
[4]u32, each non-ident source + ident control); byteid=1 throughout
(master both-broken-and-divergent → converge on the correct full
push, #263). All 111 dotbaseaddr + 3/3 784 pass; test-unit 241 green;
sizelint + smoke OK; the full w6c compiler source (214705 asm lines)
self-compiles cs==ww byte-id.
`let x: [2]inner = [inner{..}, inner{..}]` left the array unpopulated:
the N_ARRLIT per-element store handled scalar/str/float ONLY, so a
struct/array/tuple element hit the multi-word-store gap and stored just
the first 8 bytes (cs0/ww0). Both stages symmetric-broken; converge on
the populated result (#263).
Fix: an aggregate element of an array literal fills each element slot
from its source — cg_structlit_fill_bp for an N_STRUCTLIT element,
word-copy for an N_IDENT element (reusing COMMIT 2's per-element copy
shape). esz is the element's natural size (cstage esub->size). cgen.c
N_ARRLIT arm + cgenstmt.ww cglet. An aggregate `...` repeat and other
element shapes hard-stop loud (rule-7).
949 rows: arrlit_structlit, arrlit_structident (8B struct, byteid=1,
full readback). All 96 pass; test-unit 241 green; smoke OK.
The array-of-struct element store/copy family — one primitive (&(array
element) for an AGGREGATE element, used as address, never deref/truncate)
across three consumers. Both stages were symmetric-broken; converge on
the runtime-correct full-address/full-copy (#263).
(1a) `a[i].m[j] = v` (a:[N]struct) segfaulted: the `arr[i].field` arm
computed &a[i] then DEREF'd it (loaded the struct's first 8 bytes as a
value) for an `[N]T`-typed field → garbage base. Now an array-typed
field of an array element leaves the field ADDRESS (the #135 read-side,
applied to the array-element base). cgen.c arm + cgenexpr.ww cgdot
N_INDEX-lhs branch.
(1b) `a[i] = aggregateval` truncated the copy to an 8B MOVQ. New
aggregate (struct/array/tuple >8B) element-store branch word-copies the
element from the rhs source address (ident / N_DOT field / `*p` deref) —
the WRITE-twin of the #268 let-init loop. cgen.c N_INDEX store +
cgenexpr.ww cgassign.
(3a) `let c = x.arr[i]` (N_DOT base) / `let c = a[i][j]` (nested) dropped
the copy: the #268 let-init N_INDEX source-addr arm was N_IDENT-base-
gated. Now computes &base[idx] via cg_dotbase_addr (N_DOT field) or the
&abase[bidx] spine (nested N_IDENT-array base). cgen.c N_LET +
cgenstmt.ww cglet.
949 rows: elemfield_store, elem_struct_store, elem_arr_store,
letcopy_{dot,nest}_prim, letcopy_subarr (byteid=1); letcopy_{dot,nest}_
struct (byteid=0 — run-correct, byte-id blocked by the orthogonal
value-nested-struct frame divergence #254). All 94 pass; test-unit 241
green.
The fold-1b unified arm (bb2f4e1) added an N_IDENT addressable-rhs source
setup, but the two stages gated the GLOBAL case differently: cstage used
let_islet || def_isarraydef, wwstage used isletvar || deflookup (ANY def).
On a struct-typed `def` used as an aggregate-copy rhs (`let c: T = G`)
wwstage copied the whole value (correct) while cstage truncated to the 8B
scalar tail — a cs!=ww divergence (rule-10). A struct-LET global already
copies on both, so the def gap was also an internal cstage inconsistency.
Struct defs are first-class laid-out aggregates (DATA storage + field
load, #129 A.2/A.3), so converge on the correct full copy on both: add
def_isstructdef to cstage's predicate and replace wwstage's broad
deflookup with the def_is{array,struct}def pairing already held identical
in defisaddressable. 949 +2 rows (array-def + struct-def global, full
readback, byteid=1).
#265 fold-1 landed the deref-rhs aggregate copy as one slot→slot memcpy
loop fed from a source address in SI. fold-1b adds the remaining
addressable-rhs source-address setups, all routed into that SAME loop:
- array IDENT `let c: [N]T = s` — LEAQ the source slot into SI.
Pre-fix both stages truncated to the 8B scalar tail.
- N_DOT field `let c: A = o.i` — cg_dotchain_addr / dotchainaddr
(#253) lands &(o.i) in SI. Pre-fix truncated to 8B.
- N_INDEX element `let c: A = a[i]` — the &base[i] spine (#252:
scaled index + LEAQ base) lands the element address in SI. Pre-fix
scalar-loaded the element address as a value → segfault.
Size (the #254 non-slot-padded ABI extent) comes from the declared let
type for every shape (lu->size / structabisize|tinfo.size), independent
of the rhs; only the per-rhs address setup differs. The deref arm
becomes one branch of the unified arm. Struct-IDENT keeps its own #32
slot-copy arm above (unchanged). With those, the whole addressable-rhs
let-init-copy family is closed by construction: struct-ident / array-
ident / deref / N_DOT / N_INDEX all full-copy, both stages byte-identical
(rule-10).
949 gains 9 full-readback rows (every member written distinct + summed,
so a partial copy fails): array-ident 16B/32B + 12B(MOVL)/11B(MOVW+MOVB)
tails; N_DOT struct-field 16B + array-field 32B + 11B-tail struct field;
N_INDEX struct element 16B/32B. The N_INDEX source array is populated
through a `*inner` to `&a[i]` (the #135/#252 store path) because the
array-of-struct element direct store (`a[i].m[j]=v` / `a[i]=s` / struct-
array literal) segfaults on a SEPARATE pre-existing bug, reported
alongside this fold. w6c+wwdump combined.ww regen (#110). 70/70 949,
test-unit 241, sizelint, smoke green.
A `let c: T = *p` (T a struct or array, >8B) copied no full aggregate:
cstage dropped the init entirely (c read garbage); wwstage emitted only
the scalar `MOVQ AX,off(BP)` tail (first 8 bytes). Both wrong, differently
— converge BOTH stages on a size-driven slot-to-slot memcpy: cgexpr the
deref operand to the source address in AX, MOVQ AX,SI, then a MOVQ run
plus a sized MOVL/MOVW/MOVB tail over the #254 non-slot-padded ABI extent
(lu->size / structabisize for a struct, tinfo.size for an array). Mirror
arms in cgen.c N_LET and cgenstmt.ww cglet, byte-identical (rule-10).
Unblocks sha256's faithful `let copy = *h`. The by-value aggregate RETURN
ABI (array/struct return truncates to AX) is fold-2 (#267, deferred).
949 gains 6 full-readback rows (every member written distinct + summed,
so a truncated copy fails): struct{[4]u32} 16B, struct{[8]u32} 32B via
both *(&s) and *p (sha256 shape), bare [4]u32, and non-8-mult tails
([3]u32 12B → MOVL, [11]u8 11B → MOVW+MOVB). w6c+wwdump combined.ww regen
(#110). 61/61 949, test-unit 240, sizelint, smoke green.
cgreturn's passthrough predicate was TYPE-only (istagged && type-eq), with
no source-kind filter. It forwarded the source's AX/DX/CX unchanged, which
is correct ONLY when the source already materialized the full tagged slot
into registers — N_CALL / N_INDEX / N_DOT (the #261-broadened set). For a
tagged LOCAL ident, cgexpr loads only word0 (the tag) into AX, never the
payload into DX, so passthrough dropped the payload: `return v` of a
`(i32|void)=7i32` exited 0 instead of 7. wwstage was already correct — its
forwardtagged kind filter excludes N_IDENT, routing it through the
scratch-widen path. The runtime oracle (cstage 0, wwstage 7) proved cstage
is the bug; this aligns cstage UP.
Gate passthrough to {N_CALL,N_INDEX,N_DOT}; a tagged-ident return now falls
to the existing scratch-slot widen path (cg_widen_tagged_store tagged-subset
N_IDENT arm), byte-identical to wwstage's return scratch-widen. cstage-only
(no combined.ww regen — combined.ww embeds the unchanged wwstage source;
byte-id is blind here, the new 949 rows are the net).
test/949: tagged_ident_ret_i32 (7) + tagged_ident_ret_void (void tag
survives) + register-resident controls tagged_call_ret_ctrl /
tagged_dot_ret_ctrl (passthrough must still fire); INDEX control already
present. All dual-stage run + cs==ww byte-id.
#259: the tagged-union array-field indexed STORE arm computed &arr[i]
from a non-ident base (`x.o[1]=v` where o:[N](T|void)) with a plain
cgexpr(base) — the N_DOT array field auto-derefs (loads the field's
first 8 bytes AS a pointer) -> garbage dest -> SEGFAULT. Route the base
through the array-gated helper cg_dotbase_addr/dotbaseaddr (dst BX keeps
the scaled index live in AX; viaptr + chained handled by the shared
helper), mirroring #257. Symmetric both stages. This was the last
unrouted cgexpr(base) cell in the array-field-base-address family
(#135/#252/#253/#255/#257) — proof-grep of both stages now shows ZERO
unrouted base cells in the slice/decay/addr/index/store builders, so the
family is closed by construction. (The chained-ptr-field scalar/str/
float store sites at cgenexpr.ww:6489+ / cgen.c:4379+ correctly cgexpr
the pointer spine and are the #133 family, not array-field-address.)
#256: align wwstage dotchainaddr's N_IDENT non-local arm to carry
cstage cg_dotchain_addr's `let_islet || def_isstructdef` guard (here
isletvar || deflookup) instead of emitting LEAQ name(SB) unconditionally.
Unreachable on valid input (a struct-typed chain root is always local /
let-global / struct def) so zero divergent asm — never-silent ethos only.
Tests (949): store-only byte-id rows (tagged_store_own/_ptr) gate the
#259 store base-address emission cs==ww; store+readback rows
(tagged_store_*_rd) are run-only (cstage) proving the store wrote the
right slot (66/77) and no longer segfaults. byte-id on the readback rows
is blocked by an ORTHOGONAL newly-surfaced divergence in the N_DOT-base
tagged-element READ materialization (sibling of #255: wwstage loads one
word + zeroes the tag where cstage copies the full 16-byte slot) — the
store base is already byte-id; only the read-back diverges. Reported
separately for triage.
combined.ww regen'd (w6c + wwdump embed cgen).
An inline slice of a struct `[N]T`-field passed DIRECTLY as a call
argument (`rd(x.o[lo:hi])`) materialized the slice .ptr from the field
VALUE, not its ADDRESS: the pushargs/pushargsrev N_SLICE inline builder's
non-ident else-arm did plain cgexpr(base), so the N_DOT field auto-derefs
(MOVL field,AX used as .ptr) -> callee derefs garbage -> SEGFAULT. The
let-init / assign-rhs / return / hoist-to-local contexts already routed
through the cgslice #252 choke-point; only this call-arg builder kept a
private duplicate. cs==ww both segfaulted identically pre-fix (gate-blind).
Fix (symmetric both stages):
- route the else-arm through cg_dotbase_addr / dotbaseaddr (the cgslice
#252 choke-point: array-field-gated, so `[]T`/str/`*T` fields fall
through to cgexpr; chained inner `o.p.m` handled via its #253 arm);
- extend the N_IDENT-only esz gate to N_DOT bases, taking the element
width from the checker-stamped base->type (rule-13 type table), so
non-u8 call-arg slices scale stride.
Before: `MOVL -8(BP),AX; PUSHQ AX` (field value as .ptr). After:
`LEAQ -8(BP),AX; PUSHQ AX` (field address). cs==ww byte-identical.
Helper note: used dotbaseaddr (not dotchainaddr as first scoped) — it is
the established cgslice choke-point and is array-field-gated, so a slice/
str-typed field base keeps the correct cgexpr header-ptr load; bare
dotchainaddr lacks that gate and would mis-emit the field address for
those. dotbaseaddr already handles the chained `o.p.m` inner via #253.
Tests: test/wcc/949 gains 6 call-arg rows (u8, i32-esz-stride, via-*struct,
chained, + hoist-to-local and bare-local-array controls), each run-
correctness AND cs==ww byte-id.
PROOF-GREP residual: the tagged-union-element indexed-STORE arm
(cgen.c:~4972 / cgenexpr.ww:~5024) is the same N_DOT-base auto-deref shape,
still unrouted in BOTH stages (symmetric, segfaults) — a distinct
consumption axis filed separately; NOT fixed here.
cg_dotbase_addr / dotbaseaddr rejected a non-ident inner, so a chained
base (`o.p.m[i]` / `o.i.m[i]` / `o.a.b.m[i]`) fell to cgexpr(base) which
auto-derefs the array field's first 8 bytes AS a pointer -> garbage base
-> segfault (base64 fillobuf `s.enc.encmap[...]` blocker). Extend the one
helper per stage to accept a chained inner: a new cg_dotchain_addr /
dotchainaddr recovers the container base via the dot-chain spine (recurse
to &x, deref when x is a *struct, sum field offsets), keeping the same
no-AX/no-stack spill contract. dotbaseaddr then takes the pointer VALUE of
inner when viaptr, else its ADDRESS, and adds the field offset. One fix
closes every op (index r/w, addr-of, slice, compound) since all route
through the helper. Symmetric cs==ww byte-id.
test/949: +22 rows. Chained-PTR (rd/wr/addr/slice x2/compound), deeper
(value+ptr leaf links, triple-pointer exercising the internal deref),
non-u8 esz stride (i32 addr+slice), and single-level controls — all
byte-id. The chained VALUE-container arm (`o.i.m`) is run-only (byteid=0):
it needs a value nested-struct instance, which trips THREE orthogonal
pre-existing cs!=ww emission divergences (bare-let zero-init policy,
global DATAW byte count, i32 element-load opcode in the index fallback)
unrelated to #253. Run correctness proves the segfault is gone for that
cell; byte-id there awaits the separate wwstage value-nested-struct fix.
Taking &x.o[i] (address-of) or slicing x.o[lo:hi] / x.o[lo:] of a
struct's [N]T-typed FIELD computed the field's VALUE as the base
address (MOVL off(BP),AX) instead of its ADDRESS (LEAQ off(BP),AX) ->
garbage pointer -> segfault. The index read/write path was fixed in
#135; this is the unwired addr-of + slice sibling — both base-address
paths fell to the generic cgexpr(base) auto-deref.
Wire the #135 cg_dotbase_addr / dotbaseaddr helper into the addr-of
N_INDEX complex-base arm and the N_SLICE base arm, symmetric on both
stages (guarded if(!dotbase) cgexpr(base)). Extend the slice element
stride (esz) and default-hi length to an N_DOT array-field base too,
read from the field's element tinfo / array length via the type table
(rule-13) — so non-u8 element slices scale correctly and s.obuf[lo:]
gets the array's element count.
cstage already derived default-hi via base->type (alen); only wwstage
needed the N_DOT default-hi arm. cs==ww byte-identical on every shape.
test/949_dotbase_addr_slice_run: 7 dual-stage rows (addr-of local +
*struct param, explicit + default-hi u8 slice, non-u8 [4]i32 stride,
bare-local control), run + cs==ww byte-id. Regen w6c/wwdump combined.ww.
A struct literal initialising an array-typed field as a local
(`e{ encmap = [..] }`) silently dropped the initializer: cg_structlit_fill
(cstage) / cgstructlitfill (wwstage) had no TY_ARRAY field arm, so the
array field fell to the generic scalar tail — cgexpr the N_ARRLIT (→ AX≈0)
then store one sized word — losing every element. cstage returned 0;
wwstage emitted byte-identical wrong code. (The GLOBAL literal-init path
is unaffected: it goes through emit_struct_lit_bytes, already correct via
#129 A.3.)
Both stages now element-wise store the N_ARRLIT at base+field_off+i*esz,
reusing the proven N_LET array-init shape (cgen.c:8467 / cgenstmt.ww:1393)
for int and float elements plus its `...` repeat fill; esz routes through
the type table (rule 13). str/slice/struct/tagged ELEMENT arrays are the
N_LET path's documented multi-word gap (cgen.c:8462) — converted from the
silent drop to a LOUD rule-7 error in both stages, not left silent.
Symmetric both stages (rule 10), byte-identical .s.
The `...` repeat in a struct-literal array field is checker-unreachable
today (the field type-check rejects `[v...]` length inference — a
separate checker gap); the arm mirrors N_LET's repeat for symmetry.
Test 949_structlit_arrfield_run: +local literal-init reads (idx 0 / last
element), cstage run + cs==ww byte-id.
Reading an array-typed field of a module-global struct value (`G.arr[i]`)
silently miscompiled: the N_INDEX fallback's cg_dotbase_addr (cstage) /
dotbaseaddr (wwstage) helper — the #135 sibling that computes &(s.field)
for a `[N]T` field — had no module-global-struct base arm. cstage emitted
`LEAQ (BP)` (localfind returns 0 for a global, so it read the stack frame
→ 0); wwstage's localfindnode returned nil and the fallback keyed on the
FIELD name, so it returned false and the caller's cgexpr(N_DOT) loaded the
field VALUE as a pointer → SEGFAULT. The .data was already correct
(emit_struct_lit_bytes #129 A.3); only the READ base address was wrong.
Both stages now emit `LEAQ name(SB) (+ ADDQ field_off)` for a global
value-struct base, mirroring the scalar global-field read (cgen.c:7532);
const globals resolve via def_isstructdef. Symmetric both stages (rule
10), byte-identical .s. Unblocks base64's `const std_encoding.encmap[i]`
reads (#22).
Test 949_structlit_arrfield_run: global `let`/`def` struct array-field
read, cstage run + cs==ww byte-id.
cgexpr could not produce a tuple VALUE, so a destructure / let bind of an
RVALUE tuple read garbage past the first element (cstage) or left an untyped
binder aborting wwstage's asserttyped gate — a DANGEROUS gate-blind cs!=ww,
and the strconv-int blocker (Hare's stoi64/stou64 require
`let (sign, u) = parseint(s, base)?`). Three feeders, all routed at the same
SysV register-return cursor the cgmlet/cgmassign consumers already read:
- an N_TUPLE literal fell to the `cgexpr_int(0)` / `MOVQ $0, AX` default;
- a tuple-typed IDENT loaded only word0 into AX (`yield t`, `return t`,
`let q = t`), leaving DX/CX stale;
- the `?`/`!` unwrap of a tuple-in-union payload lifted only word0->AX,
stranding word1 in CX (the scalar/str success ABI).
Fix (both stages, byte-identical per rule 10):
- cgexpr packs an N_TUPLE literal into the cursor (cg_tuple_lit_to_cursor /
cgtuplelittocursor — a byte-identical reuse of cgreturn's in-register
N_TUPLE arm) and a tuple IDENT from its slot at the register-ABI stride
(cg_tuple_slot_to_cursor / cgtupleslottocursor);
- the ?/! unwrap shifts a tuple success payload down one integer reg past
the tag (cg_tagged_tuple_payload_shift / cgtaggedtuplepayloadshift),
loud-stopping a float/slice/str payload element (the SysV per-eightbyte
tagged-tuple-payload classification is #243);
- wwstage's checker recovers the popped match-arm binder type for a
`yield <binder>` operand (matchyieldtype's scope-free fallback to the
arm's declared type), so the destructured binders stamp — cstage reads
the operand's already-stamped ->type, wwstage caches only a tinfo.
Over-cap rvalue-tuple materialisation (no slot to sret a bare expression
value into) loud-stops both stages — the #10 follow-up.
NOT closed (distinct root, deferred to #238/task #6): single-var
`let q = (true, 9u64)` then `q.N` — the N_LET tuple-init sz==16||32 gate
drops a narrow-first mixed tuple, and the N_DOT tuple-field PACKED-offset
reader disagrees with tuple_store's 8B stride. Not the rvalue-into-cursor
fix and not a strconv blocker (strconv destructures); documented at the test
header.
Test 945_rvalue_tuple_destructure_run: literal destructure, match-yield
destructure, and the ?-call strconv shape, each run + cs==ww byte-id on both
drivers (9 checks). Embedded w6c/wwdump combined.ww regenerated.
A mixed-scalar tuple WRAPPED IN A TAGGED UNION (the (neg, n) shape Hare's
strconv parseint returns, ((bool,u64)|invalid|overflow)) miscompiled three
ways, all gate-blind (no bootstrap tuple-in-union):
(a) cstage CONSTRUCTION: a tuple variant fell through the N_RETURN scalar
shuffle, which ZEROED tag + payload — the operands were never packed.
Route the tuple variant through the scratch-slot widen path; add a
TY_TUPLE arm to cg_widen_tagged_store that packs each element into the
union payload at the register-ABI 8B stride + sets the variant tag.
(b) wwstage CHECKER: `let (a,b)=t` over a plain tuple ident (the match-
bound union payload) left the un-annotated binders UNTYPED, so the bin
node reading them was untyped -> asserttyped abort. The element-type
distribution only fired for an N_CALL rhs. Consume the rhs tuple type
for ANY rhs (mirror cstage check.c:2017).
(c) BOTH stages DESTRUCTURE: the register-cursor receive assumes the rhs
left every element in AX/DX/CX (a call's tuple-return ABI). For a tuple
IDENT cgexpr loads only word0->AX, so the 2nd binder read a STALE DX.
Copy each element from the ident's slot at the 8B stride.
Construction is correct at ANY variant position (the resolved tag, not a
default 0); wwstage resolves it via the typeeq core (flatvariantidxt), not
taggedvariantindext whose str/slice shape-fallback would mask a mismatch.
Two rule-7 loud-stops cover shapes this slotted packing can't yet handle,
on BOTH stages, so neither silently miscompiles:
- a tuple with a SysV-eightbyte-sharing narrow pair (e.g. (i32,i32,u64)),
caught by the 8+payload > slot-size guard (the eightbyte tuple
classification is #243);
- a tuple built from a BARE LITERAL element (`true`/`false`, suffix-less
`7`). cstage's cg_tag_for_variant can't type the literal (#241), returns
-1, and loud-stops. wwstage types `true` as bool and `7` as untyped_int,
so flatvariantidxt WOULD resolve the variant — a program cstage rejects
but wwstage accepts is the cs!=ww divergence rule 10 forbids. wwstage
mirrors cstage's CONDITION (a bare-literal element), not its -1
mechanism, with an explicit guard that aligns the richer side DOWN. Lift
BOTH guards together when #241 lands cstage literal typing -> symmetric
accept.
Test 940_tuple_in_union: 4 K_RUN rows (variant 0, void arm, tuple at
variant 1 two ways) x cstage-run + wwstage-run + cs==ww byte-id, plus 2
K_BUILDERR rows (eightbyte-share, bare-literal) asserting a loud stop with
the #242 diagnostic on BOTH drivers = 16 ok.
An over-cap tuple mixing a scalar with slices/str (e.g. (int,[]u8,str),
56B) laid out differently in the two stages — gate-blind, since no
bootstrap path returns such a tuple. Two silent cs!=ww bugs, one per
ABI side:
- callee SEND (cstage cgen.c N_RETURN over-cap-tuple arm): foff
advanced by the LITERAL expression's type size. A bare int literal
element is stamped TY_UNTYPED_INT (size 0), so `e->type->size`
added 0 for a leading scalar — the next element clobbered it at
offset 0 and every trailing element packed 8 bytes low. wwstage
already sized from the return-type tuple (c.fnret.list), so the
callee frames diverged. Fix: size foff from cg_ret_type's tuple
params (rule-13 type table), aligning cstage to wwstage and to the
t.N reader's f->offset.
- caller RECEIVE (wwstage cgenstmt.ww cglet N_TTUPLE arm): the
in-cap register tuple-receive branch had no capacity gate, so a
56B over-cap tuple was received via AX/DX/CX/R8 (+ R8 fill)
instead of from the sret dest the callee wrote. cstage gates the
twin branch on `sz == 16 || sz == 32` and falls over-cap tuples
through to the sret receive. Fix: add the same size gate to
wwstage, aligning it to cstage.
Both stages now emit byte-identical asm and the value round-trips.
Regen w6c + wwdump combined.ww (cgenstmt embeds in both).
New 940_mixed_scalar_tuple_sret_run: leading/trailing/middle scalar
shapes, annotated + inferred let, each self-asserting every element
(scalar direct, slice/str via len) — both drivers exit 0 + cs==ww
byte-id (12/12).
len() special-cased only a plain N_IDENT slice operand (load .len at
BP+off+8) and an array operand (fold $alen); every other shape fell back
to a bare cgexpr(operand), which for a slice leaves AX=.ptr. A tuple-
element read (t.N) loads only AX=.ptr, so len(t.N) on a slice/str tuple
element returned the slice's .ptr word AS its length — a silent
miscompile, gate-blind because the bootstrap never does len() on a
slice-typed tuple element (sibling of the #234/#237 tuple-sret cluster).
Both stages: detect a slice/str tuple-element len() operand and load the
element's .len word directly at BP + element_off + 8, mirroring the
N_IDENT slice arm and the tuple-field-offset walk (element_off sums
preceding element sizes through the type table). Byte-identical asm
(rule 10). The separate tuple-element-read full-header gap is #238; a
leading-scalar mixed-tuple has its own pre-existing sret-layout cs/ww
divergence, filed apart from #235.
Test 903_tuple_elem_slice_len_run: 4 slice/str-only tuple rows (two/
three slices, str+slice, slice+str; distinct lengths), build+run both
drivers + cs==ww byte-id. 12/12 ok.
The STORE-twin of the Fold-B over-cap-tuple sret RECEIVE (a937d67). Fold B
wired single-var-let / destructure / reassign / return-forward to receive a
> 4-eightbyte (sret) tuple-returning call, but a FIELD or INDEXED-lvalue
dest stayed unwired: the store dropped the callee's sret body (a truncated
MOVQ through a stale RDI) — a silent miscompile, gate-blind because the
bootstrap never field-stores a wide tuple.
Per Rob's ruling A (one class, one commit): convert the silent miscompile
into either a CORRECT store or a LOUD stop, never a fall-through.
- cstage cmd/w6c/cgen.c: the struct-field N_DOT store and the N_INDEX
lvalue store each gain an arm keyed on cg_sret_retsize(dest) > 0 &&
rhs == N_CALL. A LOCAL dest (BP-relative, not via_ptr / global) sets
cg_sret_dest_off so the callee's hidden RDI writes the WHOLE tuple
straight into the slot — field: boff + foff; indexed: boff + cidx*esz
(a CONSTANT index into a local value array, the only indexed form whose
dest is a static BP offset). Every other dest fatals "#234-tail".
- wwstage selfhost/cmd/wcc/cgenexpr.ww: symmetric (rule 10). The direct
struct-local field branch sets c.sretdestoff = lc.off + fi.foff; the
via_ptr branch, the global branch, and the N_INDEX arm hard-stop loud
with the same #234-tail diagnostic. The field branches key on
sretretsize(fi.tnode) > 0 (fi.tnode is a real type-AST node). The
N_INDEX arm keys its ENTRY on callsretsize(c, n.rhs) > 0 — the
callee-return-type SSoT (cgenutil.ww) the receive sites use — NOT on
sretretsize(elemtn): elemtn is only a type node for an N_IDENT base, a
VALUE node for an N_DOT base (`s.arr[i]`) / chained (`a[i][k]`), which
fell to sretretsize=0 and let those forms drop SILENTLY through to the
truncating store. The callee return type equals the dest-element type
(checker-guaranteed), so the verdict is byte-identical to cstage's
cg_sret_retsize, and the base-shape split then loud-stops every
non-local-array form, base-kind-independent.
Deferred (#234-tail): a via_ptr field (`p.f`), a global field (`g.f`), an
N_DOT-base index (`s.arr[i]`), a chained index (`a[i][k]`), and a runtime /
slice / pointer index all need a runtime RDI-pointer dest, which
cg_sret_dest_off (BP-relative only) can't express — they hard-error loud
(rule 7), never a truncating store.
Depends on #237 (committed first): the wwstage struct-field slot for a
tuple field is only correctly sized with that fix, so the struct-field arm
is byte-id-symmetric here.
Test 940: indexed-on-local and local-struct-field rows RUN on both stages
(exit 0) AND assert cs==ww byte-id; readback via a raw pointer
(`(&dest):*int; p[i]`) since a tuple-element read `dest.N` is a separate gap
(#238). Builderr rows assert the via_ptr / global / runtime-index /
N_DOT-base / chained-index forms loud-stop with #234-tail on BOTH drivers
(the N_DOT-base + chained rows are the regression witnesses for the wwstage
silent-store gap closed by the callsretsize re-key). The bootstrap exercises
no such store, so the w6c/wwdump combined amalgams regen with no asm change
(byte-id-neutral bootstrap; the new hard-error never fires self-compiling).
Fold A made the CALLEE emit an over-capacity tuple return (> 4 GP or > 2
SSE eightbytes) via sret, but every receive site stayed loud-stopped, so
such a fn was not yet usefully callable. Fold B wires the call/receive end
by aligning every receive gate UP to the shared cg_sret_retsize() /
callsretsize() > 0 predicate (never a kind), per Rob's (B) ruling:
- single-var-let `let t = f();` cstage gate generalised from
TY_STRUCT&&>24 to cg_sret_retsize(lt)>0; the let's slot IS the
sret dest, the callee writes the whole tuple there, t.0/t.1 read
by offset. wwstage already keyed callsretsize (verified).
- N_ASSIGN-ident `t = f();` same generalisation; global arm
kept TY_STRUCT-only (a tuple-global has no sret-to-symbol path in
either stage). wwstage grows a tuple-local arm (rettupleof gates
it apart from the >24B-struct recv, which keeps its own path).
- destructure `let (a,b) = f();` and `a,b = f();` — the genuinely
new wiring: the callee sret's into the @sretscr discard slot, then
a copy-out loop moves each element to its binding at the SAME
packed offset the SEND wrote (foff += element size), each at its
natural width (#169); a `_` binding skips its store but advances
foff. Both stages, byte-identical.
- return-forward `return f();` cstage forward gate generalised
to the predicate, reusing cg_sret_forward verbatim. wwstage
already keyed sretretsize (verified).
The escape boundary stays loud: arg-pass `g(f())` fatals identically in
both stages (tuple arg exceeds return-cursor ABI capacity).
Test 799 is the runtime net Fold A deferred (byte-id is blind to a
SEND/RECEIVE layout mismatch): the bytes.cut-shaped ([]u8,[]u8) round-trip
over destructure / single-var-let / reassign / return-forward, each both
RUN under cstage and asserted cs==ww byte-identical. Tests 945 (row F)
and 956 (f64x3) flip from asserting the old over-cap loud-stop to
asserting the now-working sret round-trip. combined.ww amalgams (w6c +
wwdump embed the wcc cgen) regenerated. Unblocks #4 bytes.cut/rcut.
A tuple return whose SysV register-return footprint exceeds the caps
(> 4 integer eightbytes or > 2 SSE eightbytes) previously LOUD-STOPPED
at the N_RETURN SEND. Fold A makes the CALLEE emit such a return through
the existing >24B-struct sret skeleton:
- classifier (cg_sret_retsize / sretretsize) grows a TY_TUPLE arm:
walk the element footprint over the SAME caps the SEND uses, and
return the tuple's natural total size (type table) when over-cap,
else 0. The gp/sse caps are factored to a single shared SSoT
(TUPLE_GPCAP / TUPLE_SSECAP — cgen.c macros in cstage, cgen.ww defs
in wwstage) consumed by the classifier AND every emit/receive site
(the SEND, the destructure guards, the cgcall arg guard) — so
classify and emit can't disagree in either stage.
- the SEND replaces the loud-stop with a write-through: cgexpr each
element, store it through *(@sretarg) at its packed layout offset
(the t.0/t.1 positional layout), each at its natural width so a
narrow tail stores MOVL/MOVB not an over-MOVQ (#169); the dest base
reloads into DX each step since a wide element clobbers AX/BX/CX.
Then the existing struct-sret epilogue (MOVQ @sretarg->AX; ret).
- the prologue already wires @sretarg when the classifier is nonzero.
The CALL/receive side is deliberately untouched: the N_MLET/N_MASSIGN
destructure loud-stops stay, so an over-cap tuple return is not yet
usefully callable. The end-to-end round-trip arrives with Fold B (#10-B).
Symmetric cstage (cmd/w6c/cgen.c) + wwstage (cgen.ww / cgenstmt.ww /
cgenutil.ww); combined.ww amalgams regenerated. Test 798 asserts the
callee now COMPILES (no loud-stop) and w6c vs w6c_ww .s byte-identical
across all-wide, str, narrow-tail, and float-over-cap shapes; no runtime
row (uncallable until Fold B). All 236 pass incl. 990-997 byte-id.
len(str-or-slice-global) was wrong in both stages, differently. cstage's
len() arm did a BP-relative slot load; localfind returns 0 for a global,
so it emitted `MOVQ 8(BP),AX` — a bogus stack slot. wwstage's arm only
handled locals; a global fell through to cgexpr, which loads the whole
header and leaves AX=.ptr, not .len.
Both stages now emit the global .len load — LEAQ name(SB),CX; MOVQ
8(CX),AX (.len field; header is ptr@0/len@8/cap@16). The LEAQ symbol
routes through the post-#1 value mangle (cstage mahint c->cur_mod,
wwstage emitsymnamehint c.curmod), not a raw name, so a private
same-module same-leaf str global can't re-open the #1 collision.
The local-str case is unchanged (control). Slice-global rows wait on
#233 (cstage rejects `let g: []u8 = [...]` init); the str global proves
the path. Byte-id-blind, so a committed runtime + cs==ww test (797) is
the net.
The cross-module dotted value-global read (`aa.v`) and addr-of (`&aa.v`)
still mangled their symbol via the non-preferring leaf lookup (cstage
masym / wwstage emitsymname), so they emitted `LEAQ main.v(SB)` — the
WRONG module's same-leaf global — returning 99 instead of 7. #1 fixed the
DATA def-site and the bare-ident load; these four dotted LOAD/addr sites
were the residual.
Thread the dotted module name (the `m` in `m.x`) — n->lhs->str /
opnd->lhs->str / lhs.str / basenm — into the existing value mangle
(cstage mahint, wwstage emitsymnamehint), the same polarity the TY_FN
branch beside each site already uses via mafn/emitfnname. The addr-of
spine-walk for a bare-root `&global.field` is a different shape and is
left untouched.
Byte-id-blind (the bootstrap has no colliding leaves), so a committed
runtime + cs==ww test (796) is the net.
A bare cross-module value-global load mis-qualified its symbol: cgen
mangled it with curmod via a non-preferring leaf lookup, so an exported
`let v` in module aa emitted both its DATA storage AND its bare-load as
main.v, colliding with main's private v. aa.getv() returned 99, not 7.
Functions were already correct (they thread a cur_mod hint via mafn /
emitfnname); value-globals did not. Both stages emitted IDENTICAL wrong
asm, so the byte-id gate was blind to it; combined.ww (frontend) is clean
-- the bug is purely in cgen. This is the cgen residual of #55 (#1 cgen
value-global module-qualifier).
Fix, symmetric in cmd/w6c/cgen.c + selfhost/cmd/wcc/{cgen,cgenexpr}.ww:
reference-site mangle uses the resolved module (curmod-prefer for bare
idents); definition/DATA-site mangle uses the decl's own module
(d->module / d.nmod) -- threaded per-site the way fns already do, via
mahint / emitsymnamehint. The fn-mangle path is left byte-for-byte
untouched.
Deviation from the signed-off spec (ratified by rob-pike after this
finding): the spec prescribed reusing the fn lookup (mod_mangle_fn /
modlookupforfn), but its first-match fallback mis-fires for value-
globals -- mod_collect export-skips exported non-fn decls (cgen.c:1059)
to keep their bare-name data ABI, so an exported leaf is absent from the
module map and the fallback grabs another module's same-leaf private
global. The value path therefore uses a distinct exact-(name,module)-or-
bare lookup (mod_lookup_value / modlookupvalue): mangle only on an exact
match, else stay bare. Byte-id-neutral on all existing single-owner code;
exported globals stay bare (ABI preserved), private stay module-qualified.
Honest boundary (rule 7): if two modules BOTH export the same value leaf,
both stay bare and the linker sees a duplicate symbol -- a correct, loud,
link-time ABI clash (like C), NOT a silent miscompile; left to the
linker, not papered over with a cgen heuristic.
Test: test/wcc/795_xmod_valglobal_run.c -- runtime (the exported global
read returns its own value, not the colliding private one) + cs==ww
byte-id, across i32-let / def-const / f64-let. Sibling to the checker
test 794_xmod_ident_prefer, which deliberately omitted byte-id because
this cgen bug diverged the asm independently.
cg_widen_tagged_store (cmd/w6c/cgen.c) and the wwstage twin cgwidentaggedstorebp (selfhost/cmd/wcc/cgenutil.ww) wrote only the tag (slot+0) and value (slot+8) in their scalar and float arms, leaving the high pad words (slot+16..sz) as stack garbage on the BP/let/assign/return-scratch path, which never pre-zeroes. A passthrough return or u8-reinterpret of a narrow scalar/float widened into a >16B union (fmt's field = (...formattable | *mods) is 32B via the str variant) then read that garbage. Both stages were wrong identically, so the byte-id gates stayed green while the runtime truncated; fmt's spread-union scalar widen is the first real consumer. Both arms now tail-zero slot+16..sz (gated size>16), mirroring the tagged-subset/struct tail-zeros and keeping the stages byte-identical (rule 10). Adds runtime test 793; regenerates w6c/wwdump combined.ww. fmt byte-id graduation still awaits the other residual, #226 (io.read nominal-remap).
Variant selection (cg_tag_for_variant / flatvariantidxt) matched union variants by exact type only, so widening a bare value (e.g. *vtable) into a union with a NAMED ptr-alias variant (stream = *vtable, in handle = (file | stream)) found no match and the tag defaulted to 0 -- the wrong variant. In the compiler this hit emitbytes' io.write(&cgoutstream.vt) once io.write took a handle, writing the asm to a garbage fd -> empty .s -> w6c_ww miscompiled everything. Add a second selection pass: when the exact pass finds no variant, structurally compare the bare source against each NAMED-alias variant's unwrapped type; exact-match still wins in pass 1 (so a bare i64 stays the i64 variant, not oserror=!i64, which kept the os/errno union building). A >=2-structural-match collision guard (extending #218's) hard-errors LOUDLY on genuine nominal ambiguity (two ptr-aliases to the same struct) instead of silently first-picking, citing #199b/#10. Symmetric across cstage (cmd/w6c/cgen.c) and wwstage (selfhost/cmd/wcc/cgenutil.ww). One-level NAMED unwrap (chained ptr-aliases unmatched, unexercised -> #17). Adds test/wcc/789 (positive widen byte-id+runtime + degenerate-ambiguity reject guard, both stages). Unblocks post-eFinal #5's handle surface. rule-10 fix-up.
Assigning a >24B by-value struct-return into a GLOBAL lvalue dropped the
struct body: the sret dest was routed to a BP scratch temp and only the
8-byte return pointer was stored (`MOVQ AX, g(SB)`); the callee wrote the full
struct to the scratch, which never reached the global. A BP-relative dest
offset cannot name a global symbol. Pre-existing GATE-BLIND silent miscompile
— both stages emit the same broken store, so byte-id (990-997) stays green
while runtime is wrong — latent until the eFinal io surface put a global
`cgoutstream: memio.stream` (>24B) on the path, where it made cgen.ww's
self-built w6c_ww buffer every function body into a corrupt global (pos stayed
0) and emit prologue-only output.
Fix, both stages, byte-identical: route the sret dest pointer to the global
symbol so the callee writes the full struct through RDI straight into the
global. cstage adds cg_sret_dest_sym, mirroring the existing str/slice global
arm (skip the @sretscr scratch, emit `LEAQ masym(sym), DI`). wwstage carries
the lhs IDENT node (sretdestnode) and emits `LEAQ name(SB), DI` via emitsymname
— identical to cstage's symbol mangling, verified cs.s==ww.s on the probe and
across 990-997. #211-family (by-value struct + global/pointer), but a distinct
site: the cstage assignment-store into a global, not the wwstage call-return.
N_LET-global static-init (`let g: T = mk()` at top level) is a separate,
independently-broken path (#221) — link-fails for init-via-call, returns 0 for
constant init — not the sret-receive gap and not on the eFinal path; deferred.
test/wcc/940_global_sret_run: global assign (plus a branched callee to defeat
const-fold), through-pointer mutation (the io vtable-callback shape that
surfaced this), and local-init/assign regressions — runtime asserts on both
stages (the net, since byte-id is gate-blind here) plus cs.s==ww.s.
Discrimination confirmed by revert+rebuild: with the global arm disabled,
global_assign emits the truncated store and exits 1.
The outer widen of a NAMED multi-variant union value into an enclosing union
mis-tagged: the store took the tagged-subset path (inner value at slot+0 plus
a sub-variant remap, collapsing every inner sub-variant onto outer tag 0),
while the match-extract reads the nested layout (outer tag at +0, inner 16B
value at +8). Store and extract disagreed, so the match selected the first
arm. Pre-existing silent miscompile, latent because error-origination sites
(`let e: io.error = <leaf>; return e`) were gate-blind — no test discriminated
a freshly-originated error at a branched caller; the io vstream surface is the
first to do so.
Fix, both stages, byte-identical: cg_variant_match (cmd/w6c/cgen.c) and its
wwstage mirror cgvariantmatch (cgenutil.ww) fall back to structural equality
of the unwrapped tagged unions when the alias collapse loses nominal identity
(a NAMED outer variant vs an unwrapped-tagged source); the widen store now
writes the inner value at slot+8 and the outer tag at +0, matching the
extract. The inner union's build/payload/extract already worked (a destructure
through the outer round-trip recovers the inner payload) — only the
outer-widen store was wrong.
Collision guard (the fallback is unsound without it): structural matching
cannot disambiguate two nominally-distinct same-shape variants in one outer
union. That is unreachable under today's nominal-lossy collapse but inverts
the moment #199b lands the nominal layer, so if >=2 outer variants
structurally match the source we hard-error at compile time citing #199b —
both stages, an enforced invariant rather than a "rare, trust it" assumption.
Folds #219: the wwstage tinfo typeeq (lib/ww/typ.ww) had no TY_TAGGED branch
and fell through to `return true` (any two tagged unions compared equal);
cstage type_eq (type.c:269) has the structural branch. The structural fallback
above is the first and only caller to compare two bare tagged unions, so #219
is unexercised — and therefore ungateable — in isolation; it folds here per
the rule-11 couldn't-split carve-out (same structural reason as #206's
N_TTUPLE fold). The added branch mirrors cstage type_eq, tightening wwstage
into alignment.
test/wcc/925_nested_union_widen_run: outer-arm select, destructure-after-
propagation (payload survives the round-trip), destructure-let, single-variant
control, and the collision-guard compile-error, each with a cstage==wwstage
byte-id check (the path is gate-blind). Interim until #199b/B-full lands the
true nominal wrapped-slot layout.
Pre-fix the N_UN TK_STAR arm applied the generic pointer-load
`MOVQ (AX), AX` to a *fn operand. cgexpr on the operand already
left AX = fn-addr (post-#180 LEAQ); the spurious second load
read the first instruction word, and the subsequent CALL AX
jumped through that junk address and segfaulted.
Cstage: cmd/w6c/cgen.c N_UN TK_STAR opens with a TY_NAMED-peel
+ TY_FN early-break — leave AX as the fn-addr cgexpr produced.
Wwstage twin in selfhost/cmd/wcc/cgenexpr.ww cgun TK_STAR walks
the n.type_ tinfo chain the same way (TY_NAMED peel then TY_FN
check) and returns before the generic load. Mirrors
ref/harec/src/check.c expr_call's STORAGE_POINTER→STORAGE_FUNCTION
path (harec skips the deref since the pointer IS the address).
Both stages must land together per rule-10 (cstage-only would
break 990-997 byte-id gates — same lesson as #180).
Probe: test/wcc/765_star_fn_deref.c, 5 rows table-driven —
minimal / branched-callee / alias-chain / fn-with-args /
fn-tuple-return. Every row is cstage-only via stage_mask
because wwstage's checker bails asserttyped on `(*f)(...)`
(filed as #181 — N_CALL type_ stamp gap on deref-call); #181's
own probe will lock the wwstage runtime once the bail lifts.
Gate-blind risk (ken's note): byte-id alone cannot catch this
class because both stages drop the SAME instruction
symmetrically, so cs.s == ww.s holds either way. Runtime
exit-code is the only correctness net here.
Combined.ww regenerated for selfhost/cmd/{w6c,wwdump}/main.
combined.ww per #110 freshness gate.
Pre-fix the N_UN TK_AMP arm fell through silently when the operand
was an N_IDENT naming a top-level function — the let/def cascade
had no TY_FN branch, so the store at the assign site picked up
whatever AX held from prior code (commonly a stale arg register).
A subsequent (*f)(...) jumped through that junk and segfaulted.
Cstage: cmd/w6c/cgen.c N_UN TK_AMP IDENT adds a TY_FN arm before
the let/def cascade, mirror of the read-arm at line 2330 — same
mafn(opnd->str, c->cur_mod) shape. Wwstage twin in selfhost/cmd/
wcc/cgenexpr.ww cgun TK_AMP IDENT uses the analogous predicate
fnretlookup(c, nm) != nil + emitfnname(c, nm, c.curmod), matching
the cstage emit on byte-id. Both stages must land together per
rule-10 (cstage-only breaks 990-997 byte-id gates).
Combined.ww regenerated for selfhost/cmd/{w6c,wwdump}/main.combined
.ww per #110 freshness gate.
Probe: test/wcc/764_amp_fn_ident.c, 6 rows table-driven —
minimal / branched-callee / alias-chain / fn-with-args / fn-tuple
-return / cross-module. Rows 1-5 gate both stages (run + .s LEAQ
check + cs.s == ww.s byte-id); row 6 cross-module is cstage-only
because wwstage bails asserttyped on `&mod.fn` (sibling project
#184, filed). Per drew option (b) the probe exercises the address
-of without (*f)(7) — deref-call runtime coverage stays with
project #181's probe once the wwstage asserttyped bail on
N_CALL(*f) is lifted.
Phase 1 cross-mod verdict = FINE for cstage (LEAQ emits via the
already-present N_DOT TK_AMP branch at cgen.c:2477-2493); WWSTAGE
fails asserttyped on the same shape → project #184.
The RETURN twin of #165: a qualifying float-struct was returned GP-only
(struct{f64,f64} in AX/DX instead of X0/X1) — value-correct via GP transport
but not SysV register-class conformant. Route each float eightbyte through the
SSE return cursor (X0/X1) and each integer eightbyte through GP (AX/DX) via
independent cursors, at the struct-return SEND and RECV, both stages, reusing
struct_float_class verbatim. Closes the temporary tuple-SSE/struct-GP
divergence opened across #164/#165.
A qualifying struct has >=1 lone f64 so maxalign is 8 and the ABI slot is an
8-multiple — no sub-8 tail — so #169's sized tail is unreachable here and the
integer eightbyte uses a full MOVQ (cstage agrees, proven by the f64i32
cs==ww byte-id). f32 / multi-float-per-eightbyte stays GP (deferred #171b);
>16B stays sret.
Gate-blind and value-correct, so the discriminator is the SEND/RECV register
class (MOVSD X0/X1 vs MOVQ AX/DX) — covered by probe 946_structret_run.
struct params were passed GP-only, so a struct{f64,f64} argument landed in
DI/SI instead of X0/X1 — value-correct for internal ww calls (the bits
round-trip) but not SysV register-class conformant. Add a per-eightbyte
classifier (struct_float_class) routing a qualifying struct's float eightbytes
through the SSE arg cursor, reusing #163's dual-cursor plumbing and #164's
field classification. A struct qualifies only when every eightbyte is
pure-integer or a lone f64 exactly filling it (and >=1 f64); anything else —
any f32, multiple floats per eightbyte, a straddling or aggregate field —
falls back to the unchanged GP path (f32 sub-eightbyte packing deferred #165b).
Both stages' predicates are alias-aware and identical in coverage.
Gate-blind and value-correct either way, so the discriminator is the callee's
receive instruction (MOVSD vs MOVQ), scoped per-function — covered by probe
946.
Tuples were unhandled as parameters — no tuple arm in arg-push, arg-pop, or
callee-recv in either stage — so a tuple param fell to the 1-GP-word else and
dropped all but its first element (integer tuple params too; floats doubly
lost). Add tuple-param arms (SEND push+pop, callee RECV) across both stages,
reusing #164's per-element SysV classify with the 6-GP (DI,SI,DX,CX,R8,R9) +
8-SSE (X0-X7) arg cursors. A frame slot @tupargscr decouples the producing
call's return cursor from the overlapping arg cursor (capture-before-clobber).
Overflow (>6 GP / >8 SSE) fails loud (rule 7). Scoped to the N_CALL producer;
first-class tuple values (ident/literal) remain a separate unimplemented gap.
Gate-blind (the bootstrap passes no tuple params) — covered by table-driven
probe 905, which proves pre-fix element-drop and the loud-stop.
A multi-float tuple return mis-routed: SEND pushed a stale AX leaving the
float stranded in X0, while RECV (#105) read every float from X0 — so a
(f64,f64) return collided both floats. Add an SSE cursor [X0,X1] parallel to
the GP cursor [AX,DX,CX,R8], placing each element by its SysV class +
within-class index (ref/qbe/amd64/sysv.c retr), symmetric send/recv across
both stages, via a generic tuple_store/tupstore+tupsse helper that #171 will
reuse for struct-return convergence. (f64,f64,f64) = 3 SSE eightbytes exceeds
the 2-register cap and now fails loud (rule 7) rather than colliding.
Unifying the 16B and 32B whole-tuple-single-var branches onto the dual cursor
was required for f64+str coexistence; it also fixes a latent str-first
single-var bug (the old 32B branch read .ptr from DX while the send placed it
in AX). No str-first or 32B tuple exists in-tree, so integer paths stay
byte-identical (990-997 green).
cstage spilled f32 args via MOVSD (8-byte); ABI-correct is MOVSS (4-byte,
single class) per SysV (ref/qbe amd64/emit.c:524 — slot-copy-through-XMM
follows the float class). wwstage already emitted MOVSS; align cstage up
via op_for(node_isf32) at the arg PUSH (cgen.c:5366) + POP (cgen.c:5469).
Byte-id-only divergence (callee reads the f32 param low-32 regardless),
but it blocked cs==ww — closes the f32-arg-push half of the float-register
family (#119/#122/#125/#157). Bootstrap-NEUTRAL (no f32-arg caller in the
990-997 gated path). Test 907_f32arg_run (f32-arg push single/multi/mixed/
stack, cs==ww byte-id). Make test 187/187 incl 990-997.
Unblocks fold-5b (strconv f32tos passes f32 to f32bits).
The N_RETURN tagged-pack scalar-variant arm did MOVQ AX,DX, but a float
variant's value is in X0 not AX -> packed stale int (broke stof64/stof32
return (f64|invalid|overflow)). Fix: float variant bridges X0->DX via a
stack slot (SUBQ $8,SP; MOVQ $0,(SP); MOVSS|MOVSD X0,(SP); MOVQ (SP),DX;
ADDQ $8,SP), gated type_isfloat/exprfloatkind. No MOVQ-xmm->gp form
exists, hence the spill (715-class, cgreturn-register-pack twin of 715's
store-to-slot). Zero-slot-first -> deterministic f32 high-4. AX-independent
-> also resolves the multi-variant cs!=ww. Bootstrap-NEUTRAL (compiler has
no float-tagged-return). Test 707 +3 rows (f64/f32/multi, slot+8 bit-exact;
f32 no-f32-arg to isolate #143). Make test 184/184 incl 990-997 byte-id.
Close A.3's deferred shape-14 (nested array). (a) emit_array_lit_bytes
gains a TY_ARRAY-element arm (mechanical clone of the TY_STRUCT-element
arm — recurses; esz=etype->size, rule-13; ...-nested loud-reject). (b)
double-index read tbl[i][j]: when the indexed element is TY_ARRAY, leave
the sub-array ADDRESS in AX instead of dereferencing (sister of #135's
N_DOT-base fix, on the N_INDEX path) — new elemisarrayc/tinfoisarray
helpers, both stages. Storage + read = one 2D-end-to-end concern (A.2/A.3
storage+LOAD precedent).
Unblocks strconv fold-4's powers_of_ten[596][2]u64 (direct double-index
access). Bootstrap-NEUTRAL (new arms gate on TY_ARRAY-element; 1D
consumers byte-identical, 990-997 green). Test 919 +2D rows + 3D +
...-nested-reject. Deferred siblings: #155 (sub-array bind / whole-
aggregate copy), #160 (global-struct-field index base).
Extract emit_array_data + emit_array_lit_bytes helpers (both stages,
mirrored) for module-level let/def with N_ARRLIT initializer or no-rhs
zero-init. Two-pass validate-then-emit: validate pass walks elements
and fails atomically on any non-foldable element (no partial-byte
emit on failure); emit pass writes element bytes after success.
Element-kind dispatch: integer via fold_int_literal byte-for-byte
preserved from pre-A.3 inline arm (bootstrap NEUTRAL — 6 live consumers
in lib/os/bufio/strings/encoding-utf8/strconv-stof_data), float via
inline bitcast + sign-XOR byte-loop (A.1 shape, no INT64_MIN — sibling
#144), struct via recursion into emit_struct_lit_bytes (A.2 helper).
Out-of-scope element kinds (ptr-elem, nested-array) rule-7 fatal.
emit_struct_lit_bytes gains TY_ARRAY field arm calling emit_array_lit_
bytes recursively — closes A.2 parked shape-15 (array-in-struct
`def D: dt = dt{tag=42, buf=[1u8,2u8,3u8,4u8]};`).
LOAD-side widened symmetric to A.2 precedent: cstage cgindex N_INDEX
direct-ident isglobal gate widened via new DefArray registry
(def_isarraydef populated in let_collect parallel to DefStruct);
wwstage cgindex N_INDEX falls through to defvartnode on letvartnode nil
(reads defent.dtnode field added in A.2). Both stages materialise
array-def via LEAQ name(SB) same as array-let.
Mid-impl rule-7 stop: refactor initially routed only rhs==N_ARRLIT
through emitarraydata, leaving nil-rhs zero-init arrays (e.g.
`let f64tos_buf: [64]u8;` in lib/strconv) silently SKIPPED → undef-ref
at link of wwstage-rebuilt selfhost binaries. Caught on first gate run
via bootstrap 994/995 RED. Fixed by adding nil-rhs branch to
emitarraydata (zero-fills arrt.size bytes) + widening wwstage caller
to route both N_ARRLIT and nil through helper. Same-class-lower-stratum
pattern (recurring across A.1 N_UN-peel, A.2 sz==8-short-circuit, A.3
nil-rhs-drop); banked as feedback memory.
Test 919 (11 rows: int-elem 1B/4B/8B + signed-N_UN-peel + float-elem
f64/f32 + def-int / def-float / struct-with-array-field shape-15 +
explicit-zero + single-elem-regression) registered. Make test:
182/182 incl. 990-997 byte-id + combined_ww_fresh.
Followups filed:
- #43 — wwstage emitletdataw str/slice-size arms lack !isarr guards;
hypothetical no-rhs [16/24]u8 triple-emits (NOT A.3-introduced;
no live consumer; 2-line parity fix)
