Two composing defects made a two-enum union mis-tag in wwstage
(live cs!=ww: `let e: (color|shape) = shape.BALL` stored tag 0 —
the color arm — while cstage stored 1). type_eq/typeeq had no
TY_ENUM arm, so ANY two enums fell into the primitive default and
compared equal; enums are nominal (harec: an enum IS its alias
type) and now compare by node identity only. Underneath, the
wwstage post-order revisit re-stamped the constant-folded enum
member (an N_INTLIT) as untyped_int, clobbering the enum stamp the
N_DOT fold applied, so the widen matcher fell to its first-variant
fallback -- the #59.9 N_BIN guard now twins on N_INTLIT (cstage
cexpr is single-pass and never clobbered).
Two parse.c operand defects. A bare `-8` (no $) double-negated to
+8: a_parsenum lets strtoll consume the sign, and the caller's
`off * sign` re-applied it (the D_INDIR arm never multiplied, which
is why -8(BP) always worked); the wwstage twin was already correct,
so this was a latent cs!=ww on hand-written source. The name(REG)
non-SB arm faked a D_INDIR off an UNCHECKED reg_lookup and silently
discarded the parsed ident and +disp -- both stages; the shape has
no encoding and now rejects loud.
The archive walk skipped every member whose header name starts
with '/', which also dropped GNU long-name members (/N) -- any
member with a >15-char filename silently never linked. Only '/ '
(symbol index) and '//' (long-name table) are non-objects; /N
members index by content (names are never consulted). A member
that fails loadimage in the selective pull also stayed silently
unlinked with rc=0; it now reports and fails the link (rule 7).
wwstage w6l diagnostics ride the same .len idiom as the w6a sweep.
Every interned symbol -- including branch labels defined via
p.label -- was emitted STB_GLOBAL, so labels were collide-able
across objects and the header's "GLOBAL symbols only" contract was
enforced nowhere (asym.isglobal was set but never read). Labels
stay interned for fixup resolution; the symtab now carries only
exported definitions, undefined externs, and reloc-referenced syms
(a reloc pre-pass marks those). rt/ensure.o drops from 15 emitted
syms to 2. Bootstrap 991-995 hold byte-identical on the new format.
Three w6a defects, one component. MOVB with an SP/BP/SI/DI register
operand encoded AH/CH/DH/BH -- emit_rex suppressed the bare 0x40
that selects SPL/BPL/SIL/DIL (the comment claimed "we always emit
REX"; encode probe: `MOVB SI,(AX)` -> `88 30` = %dh). emit_rex8
forces the byte for low-byte codes 4-7; no current w6c output emits
those shapes, so all existing objects are unchanged. LEAQ and CALL
unsupported operand shapes fell through with zero bytes and no
errs++ (every MOV arm reports); both now err. Every hardcoded
os.write diagnostic length in the wwstage was one byte short
(truncating the newline/quote); all converted to the .len idiom so
the length cannot drift again.
Two silent seams in one sweep. The for-range destructure silently
dropped the 9th+ binding in both stages (bind arrays are sized 8);
the cap now hard-stops per the DEFER_MAX/LOOP_MAX discipline (the
shape is unreachable today -- every wide-tuple construction path
already loud-stops). argtaggedwidensz, the drain-side SSoT for
widened call args, still keyed "natural push" on slot-size equality
while pushargsrev flipped to type equality with #55 -- a same-slot
subset was classified widen by push, natural by drain (counts
coincide today; a classification reader desyncs, the #48 shape).
Both sides now share the wsame key; the widen-branch tag miss for a
concrete source joins the task-2 loud-stop family (tagged subset
sources keep t=0 -- their widentag is never read, the scratch store
remaps).
Lead 17 (enum fold-failure prev+1) DISPROVED: enumvalfold gates
every member init in the checker and its fold set is op-for-op
equal to cgen's enumevalmember, so the fallback only runs inside
already-failing compiles.
Hare's rule (harec check.c binarithm): % and the bitwise/shift five
are integer-only; + - * / need numeric operands. ww grouped % with
the numeric ops, and compound assigns never op-checked at all, so
`a % b` on floats compiled half-lowered (live cs!=ww divergence),
`a %= 2.0` plain-stored the rhs (op silently dropped, both stages),
and `s += "cd"` garbled str headers. Gate both at the checker, both
stages; the cgen float-compound fallbacks and the three unknown-
compound legacy defaults (deref/global/local) demote to rule-7 hard
stops. 34 compound-on-tagged/str/slice fixtures re-pin from the old
cgen "not wired" stops to the earlier checker diagnostics; 3 new
reject fixtures pin the closed shapes.
Every variant-index lookup miss (-1) was silently clamped to tag 0
across both stages: tagged return (value + bare-void), widen store
(struct/slice/str/float/scalar arms + push fast path + field store),
widen tag-remap (identity scan + emit loop), match case compare,
is/as typetest, tryprop error remap, and the alloc-nomem propagation
(clamped to 1). All misses are checker-rejected upstream today, so
the clamps were dead -- but any future checker/cgen seam gap would
mis-tag silently (wrong arm, wrong error, false success). Rule 7:
each site now hard-stops with a per-construct diagnostic; nullable
arms keep their raw -1 by design (a miss encodes the void polarity
for `case null`). Corpus asm byte-unchanged; bootstrap fixed point
holds.
The chained value-struct walker's N_CALL arm handled only <=24B
register returns; an sret-class rhs fell to the scalar tail and
stored ONE word (the sret dest pointer) into the leaf -- silent
truncation in BOTH stages, byteid-blind. Point the callee's hidden
RDI at the BP-relative leaf slot (the single-dot #234 arm verbatim);
a ptr-root/global chain joins the #234-tail loud-stop family.
selfhost/, cmd/, internal/ join the tree-wide sweep: every section
banner dies (91 selfhost + the cmd C-style dividers -> 0); narration
and stale contracts deleted (pre-#22 bundler notes, retired
single-PT_LOAD and no-archive claims, superseded ABI tables); every
ref/harec/qbe cite, task cite, encoding/ELF contract, and rule-10
twin pointer kept; lost lifetime/rationale lines restored where the
sweep over-cut (elf_globals ownership, kwtab linear-scan). Comment-
only proven: all five wwstage tool binaries byte-identical across
the sweep; test-commit, test-byteid (161+1399, 0 pinned-divergent),
and test-bootstrap (fixed point + 991-995 byte-id) all exit 0.
The read-through banked 66 latent-bug leads (checkpoint).
Go compiles only _test.go files as tests; discovery now keys on the
_test.ww suffix alone. The line-leading-@test compatibility allowance
(noncanonical filenames admitted as test sources) is removed from both
driver stages and the coordinator. An @test declaration outside a
*_test.ww file is rejected loudly ("@test declaration outside
*_test.ww", wording byte-identical cs/ww) instead of silently running
under compose or silently dropping in a non-T build (#6). Tree audit
found zero real carriers; the two allowance fixtures flip canonical
(dep_test.ww, widget_test.ww). New pins: direnum attest-noncanon
reject row (both-stage stderr parity) and the coordinator
noncanonical_attest_rejected package row.
Go has no per-file test targets: the PACKAGE is the unit of testing
(`go test ./...`) and same-package test files compose together. The
59-target LIBRARY_TESTS fan-out (per-file -w workdirs, one Make rule
per suite) collapses to one line: `ww test -j $(JOBS) -w
out/wwbuild/wwtest-lib -I lib/ww lib/...` — the regex dir-route
precedent generalized. The lib/regex line and the per-file pattern
rule instance dissolve with it (test/lang keeps its own rule).
Measured before committing (-j4, strings edit row): old per-file
warm 4.7-4.9s; bare walk 5.2-5.6s — a real regression, so the
coordinator first gained the brief's persistent per-package workdir:
`-w DIR` on a package target forwards to wwpackage, which keys
DIR/<dir>_<pkg> per group and hands it to each inner `ww test -c`
build. Reuse stays entirely with the driver's existing
content-identity contract — the coordinator adds pure path policy,
no cache machinery. Both driver stages drop their package-target -w
rejects (forward instead); -w with -c stays rejected at the
coordinator (two ownership contracts). After: 3.9-4.0s on the edit
row, 1.0s warm no-op, 2.5s cold — faster than the old flow on every
row.
package_test's tree -w reject row becomes the positive contract
(cold+warm byte-stable stream, cs/ww same) plus the -c conflict
reject. libbyteid roster shape DECIDED: per-file fx entries stay —
every enrolled file is still standalone-buildable, so coverage is
byte-for-byte unchanged; the dir-mode entry form arrives only with
the B3 shared-helper split that first needs it. Docs: owner table,
target table, -w contract paragraph.
A module-scope let whose rhs runs code (alloc, call — peeled through
cast/?/! wrappers) emitted no DATAW slot: emit_lets' fold-fail
silently skipped the definition and every reference died at LINK
time with 'undefined reference', the one unacceptable failure mode
(rule 7). Hare's model rejects at check time (ref/harec/src/
check.c:4360 'Unable to evaluate initializer at compile time') and
routes runtime init through @init, which ww does not have — so both
frontends now reject at the declaration with identical wording.
alias_infptr_global flips compile->error as the alloc pin (its
letvartnode N_TPTR-over-N_TSTRUCT coverage lives on in the local
alias_infptr_{nest,slicecap} siblings); callinit_global_reject pins
the call shape. Corpus pin 1741/345/21/209/1166/3482.
The two loud siblings filed in d3822d77: (a) an auto-deref *[N]T
indexed base (hs[k].result) failed cgplaceaddr's N_INDEX arm, which
knew only TY_SLICE/TY_ARRAY bases — add the one-pointer hop (the
(*hs)[k] form made implicit; load shape is the slice .ptr word's);
(b) a bare slice-element aggregate source (use(sl[k])) fell off
aggarg_srcaddr's ident-array-only N_INDEX arm into the #271 loud
stop — route the miss through the same cgplaceaddr fallback the
N_DOT arm uses. Both stages, byte-identical (probe matrix: let-copy,
field-assign, call-arg, return, indexed-store all IDENT).
After this, every ADDRESSABLE aggregate source resolves: ident
(local/module-let), deref, dot-chain, and index over array/slice/
ptr-array spines recurse through cgplaceaddr; the remaining loud
paths are rvalue sources (#40 family) and the #274 const-def arg,
both documented divergences. 8 rows added to idxdot_aggcopy_test.
The last Open-driver-work bullet. `-c -o <name>` replaces the fixed
<package>.test stem for exactly one package; the coordinator rejects a
multi-package fan-out ("cannot use -o with multiple packages", Go's
`go test -o` rule) and -o without -c is rejected at the driver
("needs -c for a package target" — a plain run executes from the temp
root, so a caller-owned name has nothing to name). Both driver stages
byte-identical wording; -S keeps its single-file-only reject.
package_test gains the contract row (naming, fixed-stem absence,
artifact runs, both rejects, both drivers); the tree-mode -o row's
pinned wording follows the contract.
The dot-chain SOURCE-address walkers (cg_dotchain_addr /
dotchainaddr) deliberately have no index hop — their AX-clean spill
contract serves the chained dotbase arm — so an aggregate copy from
elem[k].field (N_DOT over N_INDEX) fell out of the walk. Every
aggarg_srcaddr consumer loud-stopped (arg push #271, aggregate
return #272, field assign, structlit fill, ident reassign) EXCEPT
the indexed element-store arm, which left the walker's result
UNCHECKED and copied through a stale SI — a silent both-stage
miscompile (out[j] = hs[k].result read garbage). Task #6.
Both stages, two twin edits each: the aggarg_srcaddr N_DOT arm falls
back to cgplaceaddr (the C4/#40 place resolver, emission-free-on-
failure so the fallback starts clean), and the indexed store site
checks the walker before the same fallback.
Still loud (filed follow-ups): auto-deref pointer-to-array indexed
bases (hs: *[4]T; hs[k].field — cgplaceaddr's index arm takes no
deref hop; the explicit (*hs)[k].field form works), and the bare
N_INDEX slice-element aggregate source.
Evaluate the call pre-grow into a per-site scratch, receiving by the
N_LET matrix (sret / float / odd-tail / GP), then grow, slot, sized
ladder. Both stages, byte-identical. The two reject pins graduate to
16B GP accept rows; three new fixtures cover 24B GP, 3B/4B tails,
16B float, and 40B sret.
25 renames (git mv, content untouched). _test.ww is what the package
coordinator's test detection and the sep loader's canonical exclusion
key on; the old *test.ww spellings survived only through the
line-leading-@test compatibility scan. Consumers updated in place:
LIBRARY_TESTS, the libbyteid roster, the 901/974/975/976 carriers that
copy or invoke these files, and the check.c/check.ww + path/ftos
comments that cite them. Closes the open-driver-work migration bullet.
A -w DIR workdir replaces the fresh .sepwork scratch with a caller-owned
persistent package-artifact tree. A package is reused only when its
freshly composed unit byte-equals the committed unit and byte copies of
the compiler/assembler recorded in the dir equal the live tools — pure
content identity, no mtimes, no hashes, every decision reproducible
with cmp against plain files. Recompiles stage at .new names and commit
by rename, unit strictly last, so an interrupted build forces a
recompile and can never leave a committed unit vouching for uncommitted
artifacts; .o/.a additionally reject zero size (ELF/ar are never
empty), while .s/.wwi accept legitimate empties (FFI-only rt). A mode
stamp pins the -T/-S shape and the artifact protocol revision. Classic
scratch keeps its exact acquire/refuse/cleanup contract; run rejects
-w; dir-mode test rejects -w; both driver stages implement identical
behavior and wording.
PROJECT_PLAN.md was a historical disclaimer wrapping five still-open
T1 driver items — those move to docs/test-system-v2.md (Open driver
work); the rest is dated 2026-08-05 design history, archived in git.
Also retired-mechanism cleanup: both drivers drop the *.combined.ww
enumeration skip (the amalgamator is gone; leftover debris now fails
loudly instead of being silently accommodated), the consumer-less
m4_combined_epoch.md5 pin is deleted, and the wwi.ww/ww.h/check.ww
comments that still described combined.ww as the live path are
re-cited to the sep reality.
x.slicefield[:] / x.strfield[2:] emitted MOVQ $0 for the omitted hi
bound on BOTH stages (N_IDENT-gated dispatch; symmetric, so byte
identity never caught it) in all four sites: cgexpr N_SLICE + pushargs
(cstage), cgslice + pushargsrev (wwstage). The new arm re-evaluates
the pure field read for its {ptr,len,cap} header and takes .len,
covering local, viaptr, dot-chain, (*p), arr[i], and global inners.
Call inners still loud-reject upstream. Sibling of the #252/#257
array-field arms.
peel-ok/sizelint-ok/primsize-ok annotations lose their tools; sites
keep the WHY in plain words. Citations of retired carriers move to
their fixture or @test successors (949_errtype_compare -> r949_*,
900_stdlib -> library owners).
Discovers test/wcc/data, hard-pins corpus identity (1,224 fixtures,
314 error / 12 compile / 136 run / 762 run-exit, sha256 name hash),
schedules both frontend cells per fixture through os.exec with
per-cell capture dirs, and emits the strict wwfix TSV result stream.
integration.sh is the black-box owner of the CLI/protocol boundary.
Scratch acquisition is one bare mkdir that refuses a pre-existing
.sepwork; the path reaches cleanup only after acquisition succeeds.
ww build and ww test -o keep the tree for the caller; ww run and
no-output ww test build under an owned /tmp/ww_{run,test}_<pid> dir
and remove it on every result, escalating rc on cleanup failure.
The content-keyed package cache is deleted from both stages; -S stops
after per-package .s emission; --sep is gone and unknown flags are
rejected; directory ww test delegates to the wwtest coordinator; the
wwstage driver routes every subprocess through os.exec.
The wwstage Makefile recipes move off the persistent <tool>_ww.d
build-in-place shape to a per-invocation mktemp workspace: the old
recipes relied on the retired pre-wipe semantics and any rebuild
would trip the new refusal on their leftover main.sepwork.
The three 989_pkgcache_* carriers lose their subject with the cache
and go with it; the rest of the wave's carrier adaptation follows in
its own commits.
Single-directory package-test planner: discovers *.ww, groups white-box
and external <pkg>_test sources, composes the combined test package,
builds it through the sibling ww driver, and runs it with the runtime's
-package/-list/-timeout-ms contract. test/package holds its E2E corpus
and the manual runtime fixture set.
A module-global base has no frame slot; treating its local-lookup miss
as offset zero wrote .ptr/.len/.cap at the caller return address. LEAQ
the symbol like the struct-field global arms do. Both stages.
The a[i] = src copy loop enumerated ident/field/deref sources; an
N_INDEX rhs fell to the scalar tail and truncated the element. Route
it through the generic place-address funnel. Both stages.
The direct-field assignment arms enumerate CALL, STRUCTLIT, and local
IDENT producers; an addressable N_DOT/N_INDEX/deref rhs fell through to
the scalar tail, so a 16-byte struct field copied only its first word.
Resolve both places through the existing address funnels and use the
tail-aware aggregate copier. Both stages.
A str `def` has no name(SB) header; the streq push arms read a frame
slot that does not exist. Load the literal through cgexpr (AX=ptr,
BX=len) instead. Both stages.
Flip the soft-default to a hard "missing package clause" error symmetrically in
both stages (cmd/wcc/parse.c + lib/ww/syntax/parse.ww): the first real decl of a
primary section with empty pathmod/resetmod and no seen clause is now rejected.
Closes the documented soft-default divergence (the 63-wrapper carve-out).
The gate flip can't be split from the migration it breaks, so this is one atomic
commit: ~80 test/wcc wrappers gain `package main;` via a shared wwtestpkg.h
helper, 6 data fixtures plus 17 asm-grep assertions update for the bare->main.<leaf>
root-helper mangle shift, and rt/ declares `package rt;` with @symbol pinning the
bare rt_ensure/rt_malloc linker names.
Root mangling narrows: the executable entry `main` stays bare (existing
carve-out), but root helper symbols become main.X. The #84 cluster is rewritten
to assert main.run distinct from aa.run/test.run; its cgen fix and bare machinery
are retained — still load-bearing for package-less module-reset deps. New
table-driven test 782_strict_package.c (6 rows, both stages).
Retiring //ww:module-reset is deferred to #24b: it is load-bearing (clears the
.wwi pathmod so the body's package clause asserts), not a vestige; fusing its
removal here would be a silent mismatch.
All byte-id gates green; full make test reports "all 335 tests passed".
Completes the #14 close-by-construction begun by the helper extraction (e7fefa3): every <=24B aggregate register-store now routes through cg_agg_reg_store/cgaggregstore and handles all tail sizes. B/C/D/G fix a silent both-stage drop of a 3/5/6/7-byte tail (byte-id-blind: both stages dropped identically, so the gate could not see it). A/F enable a previously loud-rejected shape (a cgen backend gap, not a type rejection; harec accepts and lowers it). G (global g=f() array) routes symmetrically, dest_padded=false. The #11 arr[i].f scratch loops fold into the helper (dest_padded=true, byte-id zero-change), completing the grep-audit. Pins value-assert each eightbyte (the class is byte-id-blind) and redden under each stage's independent revert; site H's 3/5/6/7 let-receive stays a loud fatal (#22).
A stack array literal returned into a tagged-union slice success variant (fn mk() ([]i32|e) = { return [10,20,30]; }) slipped past reject_arrlit_borrow — it bailed when the dst was not TY_SLICE, but a union dst is TY_TAGGED — so cgen built an all-zeros slice header: a silent both-stage miscompile (and the .ptr would dangle anyway, no outliving backing). Extend the reject to chase a TY_TAGGED dst to its slice success variant, then apply the existing reject; this is the #25/#31 treatment seen through the union, and matches Hare (rule-9). Reached by all acceptance sites (return/assign/call-arg) so the class closes by construction. Both stages converge on an identical accept/reject decision (an array literal assignable to a union is assignable to a slice or array variant; neither stage can accept it). Full support — promoting the literal to an outliving backing — is the separate #33 arc. Compile-error fixture + a positive over-reach guard (a real slice into the union still compiles). No asm emitted by a reject, so byte-id is unchanged (no floor ratchet).
cg_structlit_fill/cgstructlitfill dispatched scalar field stores as {1->MOVB, 4->MOVL, else->MOVQ} with no fsz==2 case, so a 2-byte field was stored with an 8-byte MOVQ. Interior over-stores were harmlessly overwritten by the next field, but the LAST field at the frame edge corrupted the saved base pointer: an (S|e) union success variant places the struct payload after the 8B tag, landing the last field at -4(BP), so MOVQ AX,-4(BP) wrote into saved [BP] and POPQ BP restored garbage — a silent both-stage caller-frame clobber. Route the scalar store through the existing fldstoreop/fieldstoreop helper ({1->MOVB,2->MOVW,4->MOVL,else->MOVQ}), both stages; the #13 graduation comments already pre-documented this resolution. Pure width fix, no loud-stop (scalar widths are always {1,2,4,8} and narrowing is always correct). Value-asserting pin: an i64 sentinel live across the union-maker call (detects the clobber directly) + all members, with a non-union control.
A struct/array success variant in an (S|e)! / r? unwrap dropped eightbytes on BOTH stages (byte-id blind). Two layers: (L1) the unwrap success shuffle (cgtrytaggedshift) matched no arm for a struct/array success and fell to a bare MOVQ DX,AX, materializing only w0 — widen the existing nested-TAGGED shift's gate to admit TY_STRUCT/TY_ARRAY (the in-cap union packs the payload as raw GP words past the tag, so that shift is exact); (L2) the aggregate store arms gated on rhs.kind==N_CALL and stored one word for an unwrap rhs — relax to also admit N_TRYUNW/N_TRYPROP at the three silent store shapes (arr[i]=, single-dot field, indexed-field), reusing the materialise scratch path (now #10-correct). Rule-7 LOUD-STOP for a float-bearing success variant (an SSE eightbyte cannot ride the GP {AX,DX,CX} shift, #165). The four already-loud unwrap consumers (let-receive #7, call-arg #271, assign-existing #49, resolver-field #24) stay loud; global/chained single-dot field (#16) and the sub-8-tail-through-unwrap union-maker frame clobber (#15) are separate follow-ups. Value-asserting pin, reddens under each stage's independent revert.
The in-cap aggregate-receive materialise emitted a single narrow tail MOV that fell to MOVB for a 3/5/6/7-byte sub-8 tail, storing one byte while the scratch->dest copy read the full tail from uninitialised scratch — silently dropping members at the C2c whole-element arm (arr[i]=mk()) and loud-stopping at the #11 field arm. The scratch slot is ceil-8 padded (local_alloc/localadd round to 8) and the copy reads only tsz bytes, so flipping the tail default MOVB->MOVQ stores the full register harmlessly into the slot's own pad (in-bounds for in-cap <=24B); 1/2/4-byte tails stay byte-identical. Both stages symmetric. Removes the now-redundant #11 sub-8-tail loud-stop (keeps the float #165 and over-cap #234 loud-stops). The same narrow-tail materialise recurs at 6 other cstage sites (task #14).
Retires the obsolete idx_dot_aggret_subtail_loud //ww:error fixture (both stages now compile the case) and converts it to a positive cstage run-test; the struct-field shape is byte-id-divergent only via the pre-existing #9 frame-size bug, so the value pin uses array-field shapes. Value-asserting, reddens under each stage's independent revert.
The N_TRYUNW/N_TRYPROP success shuffle materialized {ptr,len,cap} into {AX,BX,CX} only when the success variant was a str; a slice success got only MOVQ DX,AX (ptr), leaving every slice-unwrap consumer (call-arg push, let-receive store, ident-source) reading junk .len/.cap — silent on BOTH stages (byte-id blind, cstage not the oracle). Widen the success gate to type_isstr||type_isslice (cstage) / typeisstr||typeisslice (wwstage) at all four shuffle sites; str and slice share the identical 24B header shuffle. Stays str||slice-specific — a struct success variant uses a different {AX,DX,CX} ABI (task #12). Value-asserting pin (len!=cap, poison-decoy) reddens under each stage's independent revert.
Follows #6 Mech A (Fix-R); order forced (C1 first or the slice call-arg push reddens byte-id).
The arr[i].f=src legacy assign block enumerated scalar field-type arms then fell to a 1-word scalar default, so a non-call aggregate source (ident/dot/index) cgexpr'd only its first word into AX and stored one eightbyte — silent on BOTH stages (byte-id blind). The non-indexed bases (local/deref/chained/global) reach the general assign resolver's canonical aggargsrcaddr+aggcopy; the indexed arm short-circuited before it. Route the indexed base through the block's own proven &arr[i] spine into the same aggargsrcaddr+aggcopy emitters (DRY — no third copy), dual-site symmetric. Unlike #11's in-cap arm, the source is a memory address so aggcopy is a pure memcpy: float bits and the sub-8 tail transport verbatim, no loud-stop needed. Did not fall through to the general resolver because its cgplaceaddr N_INDEX arm rejects a *[N]S (TY_PTR) base (latent resolver gap, filed separately).
Contained to the indexed base + non-call aggregate-field rhs; value-asserting pins redden under each stage's independent revert.
The arr[i].f=mk() assign arm had no aggregate-field sub-arm, so a by-value aggregate field receive fell to the scalar default (one MOVQ, dropping DX/CX) — silent on BOTH stages (byte-id blind). Add a dual-site symmetric in-cap N_CALL arm mirroring C2c (c83a340): scratch-first materialise AX/DX/CX, then word-copy to (fi.foff+k*8) within &arr[i], sizing from the natural field size fi.fsz (not slotsize). Rule-7 LOUD-STOP for the three cases the in-cap GP path cannot transport: over-cap sret (#11c/#234), a float-bearing field whose eightbyte classifies SSE (#11/#165), and a 3/5/6/7-byte sub-8 tail the single narrow tail MOV cannot express (the general cascade tail is the shared C2c/#11 follow-up, task #10). Value-asserting pins (poison-seeded, redden under each stage's independent revert) plus cfail pins for the three loud-stops.
Contained to the indexed base + in-cap call rhs; arr[i].f=src (#11b) and over-cap (#11c) are separate.
cstage rejected a deref-less function-pointer call `f(21)` (f: *fn...) with
"calling non-function" while wwstage accepted it and ran correctly -- a
cs!=ww divergence. Hare auto-dereferences a pointer callee to its fn type
before the call. Peel one pointer level after the named-type chase so
cstage accepts a `*fn` callee, matching wwstage and Hare.
One level only, deliberately: the #181-cgen indirect-call path lowers the
callee VALUE as the target (CALL AX), which is the fn address for a single
`*fn` but only the address-of the fn-ptr for `**fn`. A multi-level peel
would accept `**fn` past what cgen can lower -- a silent both-stage
miscompile (returns garbage). cstage stays loud on `**fn` (rule 7);
wwstage's loop-accept of all levels (check.ww:3763) is the over-permissive
side, filed for align-down plus the deeper cgen multi-level autoderef.
Surfaced by the codegen miscompile hunt (finding C6a). Pinned by
test/lang/fnptr_derefless_call_test.ww (deref-less + canonical (*f)(21),
multi-arg, alias-chain, tuple-return; reverting the peel reds the
deref-less rows at compile).
`alloc(Outer{ x = Inner{q=10} })` dropped the nested struct-literal field:
the alloc path had its own inline fill loop with only scalar/float/str
arms, so a field whose value is itself an N_STRUCTLIT fell to the scalar
tail and stored MOVQ $0 (cgexpr leaves a whole aggregate in no register)
over the inner slot. Both stages emitted the identical wrong fill, so the
byte-id gate was blind to it.
Route alloc's fill through the existing shared structlit-fill helper (the
one the BP-relative/global/local structlit sites already use -- it handles
nested-struct recursion, N_ARRLIT, str/slice and tagged) via a new 4th
destination mode DST_PTR_SP that reloads the heap base from (SP). This
deletes alloc's divergent inline loop, the lone site lacking the recursion.
As a side effect it also fixes a latent slice-field drop in the driver's
own alloc(sepgraph{...}) (pkg.len/.cap were dropped; the consumer reads
neither -- g.n is the count SSoT). Nested-array fields are closed in-class;
a nested tuple-LITERAL field now errors loudly and symmetrically (the #49
non-addressable gap, previously dropped silently at alloc only).
Surfaced by the codegen miscompile hunt (finding C7c). Pinned by
test/lang/alloc_nested_field_test.ww (nested struct depth 1+2, nested
array, adjacent multi-nested, sibling-no-clobber; reddens on revert).
Routing preservation proven: the whole test/lang corpus is byte-identical
HEAD vs fixed except the new pin; self-compile byte-id (990-996) green.
`arr[i] = mk()` where mk returns an in-cap (<=24B) struct/tuple/array
left the result in the #4 cgreturn registers (AX/DX/CX), but the
N_ASSIGN-into-N_INDEX path had no arm for an N_CALL rhs, so it fell to
the scalar store tail: only member 0 was written and the index scale
clobbered CX. Both stages emitted byte-identical wrong code (the
documented-but-silent #31-G gap), so the byte-id gate was blind to it.
Add an in-cap N_CALL-rhs arm: materialise the return into a frame scratch
first (keeping the CALL at the frame's natural 16B alignment), resolve
&arr[i], then word-copy the full eightbyte count + sub-8 tail -- mirroring
the #4 receive shape and the #270-1b copy. The eightbyte count derives
from the element size in the type table. Over-cap returns (#234),
non-call rhs (#270-1b) and tuple literals (#121) are unaffected; the
sibling field/deref shapes stay loud (#24).
Surfaced by the codegen miscompile hunt (finding C2c). Pinned by
test/lang/idx_aggret_recv_test.ww (10 value-asserting rows: 2/3-eightbyte
structs+tuples, array elem, sub-8 tail, const/runtime index, all four
base shapes; reddens on revert).
An untyped float literal defaults to f64, so in an f32 context it was
materialized as f64 then bit-truncated by a raw MOVSS (low-32 reinterpret)
rather than narrowed -- e.g. `let x: f32 = 2.0f32; x * 3.0` multiplied by
0.0f. Twelve byte-id-gate-blind both-wrong miscompiles, all this one cause
(compare, binop, call-arg, struct-field, array-elem against an untyped
literal).
Broaden coerce_floatlit to stamp the untyped fconst type_=f32 across the
f32-context sites (assign rhs, call-arg, struct-field, array-elem) and to
descend the implicit-cast shapes (peel unary +/-/cast, recurse binop
operands AND the binop node, recurse arrlit elems), mirroring harec's
lower_implicit_cast. The existing CVTSD2SS gate then fires; cgen is
unchanged. f64 contexts are untouched -- the stamp is gated on TY_F32.
Surfaced by the float codegen sub-hunt (= the deferred #120). Pinned by
test/lang/f32_untyped_narrow_test.ww (22 value-asserting rows incl. f64
controls; reddens on revert).
`let s: str = *h` (a str/slice loaded by value through *str / *[]T)
fell through the N_UN deref arm to the scalar load, emitting a single
MOVQ that read only the 8B .ptr and left .len/.cap from stale registers,
so len(*p) returned garbage. Both stages emitted byte-identical wrong
code, so the self-compile byte-id gate was blind to it. Add a str/slice
arm that loads the full {ptr,len,cap} via cgslicehdr when the chased
pointee is TY_STR/TY_SLICE.
Surfaced by the codegen miscompile hunt (finding C1b). Pinned by
test/lang/deref_hdr_test.ww, which interposes a different-sized decoy
header so the test reddens when the arm is reverted.
Reading gp.f through a module-global pointer miscompiled in BOTH stages,
differently: cstage classified gp as a local at boff 0 and derefed BP
(MOVQ (BP),BX), wwstage collapsed gp.f to an undefined global symbol f
(MOVQ f(SB)). Both now load the pointer value from the global's data slot
before the field offset, converging on MOVQ gp(SB),BX; MOVQ off(BX),AX.
cstage mirrors the #6 store decline; wwstage gains a global-ptr arm and
shares a cgptrfieldload helper with the local arm.
Fused, not split: the two stages must emit byte-identical asm, so a
one-stage commit would fail the byte-id gate. Sibling byte-divergences
filed: #16 (chained-spine gp.x.y), #17 (>32B tagged word-order).
Test: table-driven 689_globptr_field_read_run (24 rows, runtime + byte-id).
A module-global pointer's field store/compound (`let gp:*S=nil; gp.f += 5`,
`gp.in = Inner{...}`) SEGV'd in cstage: the enumerated N_DOT-lhs arms load
the base pointer with `MOVQ boff(BP),BX`, valid only for a LOCAL ptr slot,
but a module-global ptr has no local slot (localfind=0) so it dereferenced
the saved BP. wwstage was correct -- it routes these through its F6
cgplaceaddr resolver (its dedicated arm is scalar-`=`-only by design,
#60/#61). The byte-id gate was blind (no global-ptr compound in the
bootstrap corpus) and the deferral note was stale: this is a live cs!=ww
divergence with wwstage as the oracle.
cstage already has an equivalent assign-resolver (cgen.c ~7488) that emits
byte-identically to wwstage's F6 route, but the enumerated arms intercepted
the global case first. Fix (align cstage UP, cstage-only): two precondition
entry-guards decline a module-global `*struct` base for the compound +
non-scalar-field cases so they fall through to the resolver. Plain-scalar
`=` stays in the enumerated arm (its #47 fix already matches wwstage). The
decline and resolver accept-sets exactly partition the global-base
N_DOT-lhs space (no gap, no overlap); tagged/float field stores now both
loud-stop symmetrically (were SEGV'ing). The discriminant keys on
localfind-presence + let_islet, so a param at offset 0 stays local.
New both-stage + byte-id test 689_globptr_field_store_run covers offset-0/8,
compound, struct/str field, chained gp.x.y, indexed gp.a[i].f, with local +
offset-0-param controls. The field-READ path is independently broken in
both stages (filed #15). make clean && make test: all 403 passed, byte-id
990-996 green.
