Compare commits

...

1 Commits

Author SHA1 Message Date
coah
f67d0d010f decomp: restore Korean scan entry and lookup widths 2026-10-02 19:50:29 -05:00
3 changed files with 135 additions and 6 deletions

View File

@@ -11,15 +11,15 @@ typedef struct {
ziU8 characterLow;
} ZiKoreanKeyEntry;
ziU32 Zi8_8148302C(ziU16 key, ziU8* table ZI_NEED_WORK) {
ziU16 Zi8_8148302C(ziU16 key, ziU8* table ZI_NEED_WORK) {
ziU32 value;
ziU16 count;
ziU32 i;
ziU16 i;
count = Zi8GetTableCount(ZI8_LANG_KO, 9, ZI_WORK);
for (i = 0; (ziU16)i < count; i++) {
value = (ziU16)(((ziU16)table[((i & 0xFFFF) << 3) + (i & 0xFFFF) + 7] << 8) |
table[((i & 0xFFFF) << 3) + (i & 0xFFFF) + 8]);
for (i = 0; i < count; i++) {
value = (ziU16)(((ziU16)table[(i << 3) + i + 7] << 8) |
table[(i << 3) + i + 8]);
if (value == key) {
Zi8LogError(0x64, ZI_WORK);
return i;
@@ -313,7 +313,8 @@ ziU32 Zi8GetKOcandidates(ziGetParam* param, ZiKoreanCandidateOptions* options ZI
param->candidates[0] = 0;
}
Zi8LogError(100, ZI_WORK);
if (param->wordCharCount != 0) {
if (param->wordCharCount == 0) goto scan_key_table;
{
if (((search.keyIndex = Zi8_8148302C(param->currentWord[0], keyTable, ZI_WORK)) != 0xFFFF) &&
((((ZiKoreanKeyEntry*)(((search.keyIndex << 3) + search.keyIndex) + keyTable))->keyBytes[4] & 8) != 0)) {
search.wordOffset = ((ZiKoreanKeyEntry*)(((search.keyIndex << 3) + search.keyIndex) + keyTable))->childLow |
@@ -438,6 +439,7 @@ search_table:
param->count = search.candidateCount;
param->letters = search.candidateCount;
}
scan_key_table:
if (param->elementCount > 9) {
param->letters = 0;
return 0;

View File

@@ -0,0 +1,51 @@
# Korean empty-context flow recovery, 2026-10-03
Base: 1fa33882; branch agent/bittle/korean-candidate-reconstruction.
Owned source: zkokeyp.c / Zi8GetKOcandidates only.
## Target evidence
The first context-length test at target function+0x158 sends an empty current
word directly to+0x730, the input-length/key-table scan setup. The compiled
source instead sent this case to its+0x6C8, the prefix-search result copy and
candidate cleanup. That block corresponds to target+0x6BC, which is reached
only after a nonempty context lookup attempt or word-table search.
Added an explicit empty-context guard to a label immediately before the
existing input-length check. The nonempty prefix-search paths still enter
search_table and perform the original result/candidate cleanup. The change
recovers the target control-flow edge without moving any body or introducing
storage, helper calls, or artificial register constraints. The old edge did
not necessarily change visible output for initialized empty context; this
is a target-flow correction, not a claim of a newly demonstrated user bug.
## Validation
Fresh baseline: /tmp/ezi-korean-recon-baseline.json.
Fresh final report: /tmp/ezi-korean-flow-full-report.json.
Full build: /tmp/ezi-korean-flow-full-build.log.
Detailed gates: /tmp/ezi-korean-flow-verification.txt.
Zi8GetKOcandidates: 97.96712% ->97.97459%.
Unit fuzzy: 98.65385% ->98.65769%.
Exact functions remain1/7, exact code332/5200, matched data96/180.
All1027 units were checked: the unit fuzzy gain is the only changed measure,
with no code/data/function or fuzzy regressions. No new exact function claimed.
The source object differs from its preserved baseline at exactly one .text
payload byte, the corrected branch displacement. All section sizes and all
canonical relocation tuples remain unchanged. .data40, extab56 and
extabindex84 are byte-identical to baseline. The unit's existing exact
Zi8_81483118 remains83/83 instructions and ctxdiff diffs0, with its exact
10-entry generated switch table unchanged.
Pool first: identical and empty. Main still673/669 instructions; the known
extra zero materialization, two lookup-result narrowings and return-value
move remain. All92 mapped direct main branches now preserve their target
destinations; the17 ordered helper calls agree. The filter's43 direct branches
also agree. These are advisory correspondence checks, not an exactness claim.
Full43U build passes with the approved wrapper and WIBO_SJIS_MISSING_IMPORTS=1.
DOL SHA1: 26116613f624061ba99c8d1a299aaa6efa85670d.
git diff --check passes. No header, metadata, linking, remote, assembly,
volatile, dummy-storage or forced-register changes.

View File

@@ -0,0 +1,76 @@
# Korean lookup halfword contract, 2026-10-03
Parent: dea3851e; its empty-context correction is unchanged.
Owned source: zkokeyp.c / Zi8_8148302C and its two local callers.
## Recovered type boundary
The lookup traverses a table whose count is ziU16, comparing and addressing
records through a halfword index, and returns either that index or0xFFFF.
For countN<=65535, successful indexes are0..N-1. The loop exits atN; a
halfword index cannot wrap on any traversed iteration. There is no need for
a full-word index or return value in this contract.
Both callers store the result into the halfword search.keyIndex. The target
stores the returned register directly through the halfword store. The first
caller then narrows for its assignment-value comparison, while the second
reloads the halfword for comparison. The current full-word declaration made
MWCC emit an additional narrowing before both stores.
Recovered the helper's ziU16 return and ziU16 loop index together, then used
the typed index directly in the loop condition and nine-byte address
expressions. The old explicit masks were the reconstruction of those same
halfword promotions while the local had the wrong full-word type. They are
not stripped from a full-word value or replaced by register constraints.
The resulting helper body is byte-for-byte identical to its previous body,
including its59 instructions and resolved calls. This demonstrates that the
same lookup operations and return representation are preserved while its
callers recover the target halfword assignment boundary. Both source callers
are in this unit. A scan of all extracted target objects found the same two
text references and only the helper's own extabindex reference otherwise.
No shared declaration or header changes are needed.
## Separate trials
A native unsigned-int return alone and explicit caller-store casts were
codegen-neutral and were not retained. The coherent halfword index/return
pair is the sole selected candidate. No old Chinese ABI experiment or
arbitrary declaration/register sweep was repeated.
## Validation
Baseline: /tmp/ezi-korean-flow-full-report.json.
Fresh full report: /tmp/ezi-korean-lookup-full-report.json.
Full build: /tmp/ezi-korean-lookup-full-build.log.
Detailed gates: /tmp/ezi-korean-lookup-verification.txt.
Zi8GetKOcandidates: 97.97459% ->98.146484%.
Unit fuzzy: 98.65769% ->98.746155%.
Relative to initial1fa33882, unit98.65385% ->98.746155%.
Exact functions remain1/7, exact code332/5200, matched data96/180.
All1027 unit measures and all per-function fuzzy scores were checked: the
Korean unit/main-function gains are the only metric changes, with no
regressions. No new exact function or complete unit is claimed.
Every other function's disassembly is baseline-identical, including the
changed lookup helper. The existing exact packing helper remains83/83
instructions and ctxdiff diffs0. Main shrinks673 ->671 instructions versus
target669, from0xA84 to0xA7C versus target0xA74. Its0x60 frame is preserved.
The remaining normalized differences are one zero materialization and one
call-result move; no further width or flow difference is identified.
.data40 and extab56 retain identical payloads and canonical relocations.
The already-unmatched84-byte extabindex changes only its main-function size
byte at0x4F,0x84 ->0x7C; its score remains97.61904%, and exact-data accounting
is unchanged. No real data object or section extent was changed.
Pool first: identical and empty. All92 mapped main branches preserve their
corrected target destinations; all17 helper calls correspond. An associative,
register-value-aware advisory audit pairs all163 store/call/branch effects
without differences. These checks are not formal equivalence or exactness
claims. Full43U build passes with the approved wrapper and
WIBO_SJIS_MISSING_IMPORTS=1. DOL SHA1:
26116613f624061ba99c8d1a299aaa6efa85670d.
git diff --check passes. No linking, remote, header, assembly, volatile,
dummy-storage, undefined-value or forced-register change.