Compare commits

...

1 Commits

Author SHA1 Message Date
coah
0dfb3a551e decomp: reconstruct alpha output and dictionary scan 2026-10-02 17:26:54 -05:00
2 changed files with 76 additions and 18 deletions

View File

@@ -920,8 +920,7 @@ selectDictionary:
if ((((wordLength == 0) && (*punctuationCursor != 0)) && (dictionaryExact != 0)) &&
((((ZiAlphaWork*)workData)->singleCharacter != 0 && (parameters->elementCount == 2)))) {
*wordCursor = ((ZiAlphaWork*)workData)->singleCharacter;
wordCursor[1] = *punctuationCursor;
punctuationCursor = punctuationCursor + 1;
wordCursor[1] = *punctuationCursor++;
wordCursor[2] = 0;
wordLength = 2;
}
@@ -1208,10 +1207,8 @@ emitCandidate:
if (((ZiAlphaOptions*)optionData)->maxResults <= candidateCount) goto finishCandidates;
} else {
if (((ZiAlphaOptions*)optionData)->lookupMode != '\0') {
for (index = 0;
(index <= (int)wordLength &&
(wordCursor[index] == ((ZiAlphaOptions*)optionData)->dictionary[index]));
index = index + 1) {
for (index = 0; index <= wordLength; index++) {
if (wordCursor[index] != ((ZiAlphaOptions*)optionData)->dictionary[index]) break;
}
if ((((((ZiAlphaWork*)workData)->operation != '\0') || ((int)wordLength <= index))
&& ((((ZiAlphaWork*)workData)->operation == '\0' ||
@@ -1382,21 +1379,17 @@ finishDictionaryPass:;
}
if (((candidateCount == 0) && (punctuationCandidate)) && (firstCandidate == 0)) {
candidateCount = 1;
if ((((ZiAlphaOptions*)optionData)->countOnly == '\0') && ((parameters->getOptions & 0xfd) != 0x81)) {
*encodedOutput = 0x3e;
encodedOutput[2] = 0;
encodedOutput[1] = 0;
}
else {
if (((ZiAlphaOptions*)optionData)->countOnly != 0 || (parameters->getOptions & 0xFD) == 0x81) {
if (parameters->getMode == 1) {
*wordCursor = *elements;
} else {
*wordCursor = 0x3E;
}
else {
*wordCursor = 0x3e;
}
wordCursor[2] = 0;
wordCursor[1] = 0;
wordCursor = wordCursor + 2;
wordCursor[1] = wordCursor[2] = 0;
wordCursor += 2;
} else {
*encodedOutput = 0x3E;
encodedOutput[1] = encodedOutput[2] = 0;
}
if (((ZiAlphaOptions*)optionData)->lookupMode == '\0') {
((ZiAlphaWork*)workData)->singleCharacter = 0;

View File

@@ -0,0 +1,65 @@
# Alpha output and scan reconstruction follow-on, 2026-10-02
Parent candidate: `94b7f57e`. Its source/log were left unchanged during
independent side experiments. No ABI/header/linking changes are included.
Changed source: `libs/RVLMiddleware/eZiText/src/clib/zi8alpha.c` only.
## Target evidence and retained changes
- At `Zi8AlphaGetCandidates +0x3178`, the original emits wide-character
output before the encoded-byte alternative. The wide path assigns the
same zero to output elements 2 and 1 (+0x31A4/+0x31A8), then advances by
two characters. The encoded path likewise shares its zero assignment.
Restored that positive branch and ordinary chained zero assignments.
- The dictionary comparison loop branches from +0x2904 to its bound check
at +0x292C. The comparison body exits on a character mismatch at +0x2924;
a matched character increments the index at +0x2928. Replaced the merged
`for` condition with the corresponding bounded loop and explicit break.
- The punctuation output block at +0x1B0C..+0x1B1C reads the current cursor,
emits its character and advances that same cursor. Expressed this with
`wordCursor[1] = *punctuationCursor++`. It does not reuse a mutated index
in another operand or introduce an unsequenced read/write.
## Independently compiled, isolated trials
Each trial used the exact 43U MWCC flags from the live Ninja command and a
separate object/report under `/tmp/ezi-side`, without altering the reviewed
worktree candidate. Report columns below are function fuzzy, unit fuzzy,
matched data bytes and exact-function count:
- Early candidate-skip block: 94.65585, 96.10635, 408, 11; excluded
- Wide output first: 94.887985, 96.27548, 516, 11; retained
- Next-dictionary switches: 94.4037, 95.92264, 408, 11; excluded
- Explicit dictionary-scan break: 94.62088, 96.08087, 516, 11; retained
- Combined candidate-count limit: 94.35884, 95.88995, 408, 11; excluded
- Punctuation cursor increment: 94.4782, 95.97692, 516, 11; retained
- Wide output plus scan: 95.02306, 96.37389, 516, 11
- Final three-change subset: 95.02813, 96.37759, 516, 11
## Full current-worktree validation
Fresh report: `/tmp/ezi-alpha-followup.json`.
Full build log: `/tmp/ezi-alpha-followup-build.log`.
Exact functions 11/12 unchanged; exact code 5880/21664 unchanged;
exact data 516/564 unchanged. Unit fuzzy 95.97323% -> 96.37759%.
Function fuzzy 94.47314% -> 95.02813%, about 88 additional weighted bytes.
Relative to initial 449529d9, unit fuzzy improves 94.17892% -> 96.37759%.
Pool first: identical (no narrow strings in either pool).
Main function: 3946/3946 instructions, 0x3DA8 bytes each, frame 0x2B0.
Strict ctxdiff remains nonzero (3745 differences), dominated by the known
scalar-stack displacement, instruction placement and GPR allocation. No
new exact function is claimed. Each of the existing 11 exact functions
retains identical instruction counts and ctxdiff diffs 0.
All non-text section scores match the parent: .rodata, extab and extabindex
remain 100%; the existing 48-byte generated switch table remains unmatched.
The advisory register-normalized/control-flow alignment checked 701 branch
pairs with zero destination mismatches and 73/73 helper calls. This is a
coverage check, not an exactness claim.
Full 43U build succeeds with the approved wrapper and
`WIBO_SJIS_MISSING_IMPORTS=1`. DOL SHA1:
`26116613f624061ba99c8d1a299aaa6efa85670d`.
`git diff --check` succeeds. No new assembly, volatile workaround, register
forcing, undefined value, dummy storage, prototype or shared-header change.