diff --git a/libs/RVLMiddleware/eZiText/src/clib/zi81key.c b/libs/RVLMiddleware/eZiText/src/clib/zi81key.c index 3219959c..9e35dba6 100644 --- a/libs/RVLMiddleware/eZiText/src/clib/zi81key.c +++ b/libs/RVLMiddleware/eZiText/src/clib/zi81key.c @@ -311,21 +311,21 @@ Zi8UInt Zi8SpellingPY(ziU16 *output,Zi8UInt key,ziU8 includeTone) return length; } -ziU8 Zi8IsMatch1Key(ziU16 *input,Zi8UInt inputLength,Zi8UInt key,ziU8 requireFull,ziU8 usePinyin,ziU8 tone) +ziU8 Zi8IsMatch1Key(ziU16 *input,Zi8UInt inputLength,ziU16 key,ziU8 requireFull,ziU8 usePinyin,ziU8 tone) { ziU16 spelling[16]; Zi8UInt spellingLength; - if ((key & 0xffff) == 0) { + if (key == 0) { return 0; } if ((inputLength & 0xff) == 0) { return 1; } if (usePinyin != 0) { - spellingLength = Zi8SpellingPY(spelling,key & 0xffff,tone & 0xff); + spellingLength = Zi8SpellingPY(spelling,key,tone & 0xff); } else { - spellingLength = Zi8SpellingZY(spelling,key & 0xffff,tone & 0xff); + spellingLength = Zi8SpellingZY(spelling,key,tone & 0xff); } if ((requireFull != 0) && ((inputLength & 0xff) != (spellingLength & 0xff))) { return 0; @@ -1327,7 +1327,7 @@ int Zi8Get1KeyPressCandidates(Zi8OneKeyParam *params,Zi8OneKeyOptions *options,Z } if (elementIndex == 0) continue; if (!emitWords) { - if (trackDuplicates && Zi8SetFindCand(params->scratch,candidateOrdinal,work) != 0) continue; + if (trackDuplicates && (ziU8)Zi8SetFindCand(params->scratch,candidateOrdinal,work) != 0) continue; for (index = 0; index < candidateCount; index++) { if (candidateCharacter == output[index]) break; } @@ -1396,7 +1396,7 @@ int Zi8Get1KeyPressCandidates(Zi8OneKeyParam *params,Zi8OneKeyOptions *options,Z } if (elementIndex == 0) continue; if (!emitWords) { - if (trackDuplicates && Zi8SetFindCand(params->scratch,candidateOrdinal,work) != 0) continue; + if (trackDuplicates && (ziU8)Zi8SetFindCand(params->scratch,candidateOrdinal,work) != 0) continue; for (index = 0; index < candidateCount; index++) { if (candidateCharacter == output[index]) break; } @@ -1519,7 +1519,7 @@ int Zi8Get1KeyPressCandidates(Zi8OneKeyParam *params,Zi8OneKeyOptions *options,Z } if (elementIndex == 0) goto nextPhrase; if (!emitWords) { - if (trackDuplicates && Zi8SetFindCand(params->scratch,phraseCharacter,work) != 0) goto nextPhrase; + if (trackDuplicates && (ziU8)Zi8SetFindCand(params->scratch,phraseCharacter,work) != 0) goto nextPhrase; for (index = 0; index < candidateCount; index++) { if (firstOrdinal == output[index]) goto nextPhrase; } @@ -1601,7 +1601,7 @@ finishPhrases: } if (elementIndex == 0) goto nextSortedOrdinal; if (!emitWords) { - if (trackDuplicates && Zi8SetFindCand(params->scratch,ordinal,work) != 0) goto nextSortedOrdinal; + if (trackDuplicates && (ziU8)Zi8SetFindCand(params->scratch,ordinal,work) != 0) goto nextSortedOrdinal; ordinal = Zi8Ord2Uni(ordinal,work); for (index = 0; index < candidateCount; index++) { if (ordinal == output[index]) break; @@ -1629,11 +1629,11 @@ nextSortedOrdinal: ; } } - if ((options->countOnly == '\0') && Zi8GetZHuwdPtr(&userEntries,&userCount,work) != 0) + if ((options->countOnly == '\0') && (ziU8)Zi8GetZHuwdPtr(&userEntries,&userCount,work) != 0) { for (phraseOrdinal = 0; phraseOrdinal < userCount;) { if ((params->elementCount == '\0') || ((userEntries[1] & 0x80) != 0)) { - userOrdinal = (userEntries[1] & 0x7f) << 8 | (ziU16)userEntries[2]; + userOrdinal = (ziU16)(userEntries[1] & 0x7f) << 8 | (ziU16)userEntries[2]; phraseTable = syllableTable + (Zi8UInt)userOrdinal * 0xc; if (((*phraseTable & languageMask) != 0) && ((characterSetTable == 0 || ((characterSet & characterSetTable[(Zi8UInt)userOrdinal]) != 0)))) { @@ -1654,7 +1654,7 @@ nextSortedOrdinal: } if (elementIndex == 0) goto nextUserOrdinal; if (!emitWords) { - if (trackDuplicates && Zi8SetFindCand(params->scratch,userOrdinal,work) != 0) goto nextUserOrdinal; + if (trackDuplicates && (ziU8)Zi8SetFindCand(params->scratch,userOrdinal,work) != 0) goto nextUserOrdinal; userCharacter = Zi8Ord2Uni(userOrdinal,work); for (duplicateIndex = 0; duplicateIndex < candidateCount; duplicateIndex++) { if (userCharacter == output[duplicateIndex]) break; @@ -1710,7 +1710,7 @@ nextUserOrdinal: if (unicodeCharacter == previousCharacter) goto nextOrdinal; previousCharacter = unicodeCharacter; if (!emitWords) { - if (trackDuplicates && Zi8SetFindCand(params->scratch,ordinal,work) != 0) goto nextOrdinal; + if (trackDuplicates && (ziU8)Zi8SetFindCand(params->scratch,ordinal,work) != 0) goto nextOrdinal; for (index = 0; index < candidateCount; index++) { if (unicodeCharacter == output[index]) break; } @@ -1738,8 +1738,7 @@ nextOrdinal: } finishCandidates: if ((emitWords) && (outputIndex != 0)) { - output[outputIndex] = 0; - outputIndex++; + output[outputIndex++] = 0; } params->count = candidateCount; Zi8LogError(100,work); diff --git a/tools/decomp-assist/allhands-ezitext-onekey-boundaries.attempts.md b/tools/decomp-assist/allhands-ezitext-onekey-boundaries.attempts.md new file mode 100644 index 00000000..9c0b1a05 --- /dev/null +++ b/tools/decomp-assist/allhands-ezitext-onekey-boundaries.attempts.md @@ -0,0 +1,108 @@ +# One-key candidate boundary reconstruction, 2026-10-02 + +Parent commits: alpha 94b7f57e and 35677806, neutral Chinese ABI 105f3b45. +This pass changes only `libs/RVLMiddleware/eZiText/src/clib/zi81key.c`. +Alpha source/logs and Chinese source/log are unchanged. + +## Retained target-defined boundaries + +- `Zi8IsMatch1Key` consumes a halfword key. Six candidate calls narrow that + argument after Zi8GetPCode: target +0x0B0C, +0x0F30, +0x16C0, +0x1BA4, + +0x1F4C and +0x22A8. Restored the ziU16 formal and removed its three + redundant key masks. The helper itself remains 97/97 instructions, + objdiff 100%, ctxdiff diffs 0. All source call sites are in this same unit. +- Each of the six duplicate-bit checks consumes a byte result in the target. + Made the byte conversion explicit at these call boundaries rather than + changing the already-exact full-width duplicate helper. +- The user-dictionary availability result is narrowed to a byte at target + +0x1E00. The existing full-width declaration remains; the caller now + performs the demonstrated conversion explicitly. +- User ordinal decoding narrows the masked high byte to a halfword before + its shift at target +0x1E44. Recovered that type boundary. +- Final output termination stores through the current output index and then + increments that index without reloading it (+0x2524..+0x2534). Recovered + the ordinary postfix-index store. Its right-hand side is constant zero, + so there is no unsequenced index read/write. + +These are source/ABI shape reconstructions with demonstrated target +conversions; no callee-saved-register or declaration-order sweep was used. + +## Isolated source trials + +Each trial used the exact 43U Ninja compiler command in `/tmp/ezi-onekey-pass`, +without modifying the worktree until the final candidate was selected. +Below: candidate-function fuzzy / unit fuzzy / matched data / exact functions. + +- Halfword formal retaining old masks: 95.26784 / 97.65741 / 1232 / 4; rejected +- Byte duplicate-helper return: 94.73383 / 97.4822 / 1232 / 5; rejected +- Explicit duplicate-result bytes alone: same as preceding; not retained alone +- User-dictionary result byte alone: 95.240715 / 97.708664 / 1280 / 5 +- User ordinal halfword alone: 95.30121 / 97.735695 / 1280 / 5 +- Combined count-limit increments: 95.106384 / 97.64865 / 1232 / 5; rejected +- Terminal postfix index alone: 95.4368 / 97.79627 / 1280 / 5 +- Halfword formal/old masks + duplicate bytes + limits: + 95.984566 / 97.97763 / 1232 / 4; rejected +- Halfword formal with ordinary typed uses: + 95.26784 / 97.72078 / 1232 / 5; restores the exact helper +- Typed key + duplicate bytes: 95.83312 / 97.97334 / 1280 / 5 +- Typed key + duplicate bytes + count limits: + 95.984566 / 98.04101 / 1232 / 5; excluded for data regression +- Preceding plus terminal postfix: 96.01794 / 98.055916 / 1232 / 5; excluded +- Preceding plus user-dictionary/ordinal conversions: + 95.926155 / 98.01491 / 1232 / 5; excluded +- Typed key + duplicate bytes + terminal postfix: + 95.858154 / 97.98453 / 1280 / 5 +- Typed key + duplicate bytes + user-dictionary byte: + 95.87484 / 97.99198 / 1280 / 5 +- Typed key + duplicate bytes + ordinal halfword: + 95.93533 / 98.01901 / 1280 / 5 +- Typed key + duplicate bytes + terminal postfix + user-dictionary byte: + 95.86858 / 97.98919 / 1280 / 5 +- Typed key + duplicate bytes + terminal postfix + ordinal halfword: + 95.98957 / 98.04324 / 1280 / 5 +- Final complete boundary subset: 96.41927 / 98.23523 / 1280 / 5 + +After selecting that subset, the three count-limit blocks were tested +individually against it. OEM: 96.17564 / 98.12637 / 1280 / 5; PUD: +95.725075 / 97.92507 / 1232 / 5; phrase: 95.93575 / 98.019196 / 1280 / 5. +All three were rejected. No count-limit rewrite remains in the candidate. + +## Full validation and exact data disclosure + +Fresh report `/tmp/ezi-onekey-boundary.json`; full-build log +`/tmp/ezi-onekey-boundary-build.log`. + +Unit fuzzy 97.778564% -> 98.23523%. Candidate-function fuzzy +95.39716% -> 96.41927%, about 98 fewer weighted deficit bytes. +Exact functions remain 5/9; exact code remains 3952/21460; +matched data remains 1280/1388. All five baseline exact functions have +ctxdiff diffs 0 and identical instruction counts. + +Candidate instructions: 2387 -> 2405 versus target 2397; size +0x254C -> 0x2594 versus target 0x2574. Frame remains 0x140. +It is not exact, and no new exact function is claimed. + +The spelling tables remain 1152 byte-identical bytes. .sdata2 and extab +remain byte-identical at 8 and 72 bytes. The switch table's source is 44 +bytes versus target 48, the pre-existing four-byte section-alignment tail; +all 44 bytes and all eleven canonical relocation targets/addends match. +No real data object was resized or renamed and no padding was added. + +The already-unmatched 108-byte extabindex still has two differing bytes at +0x4F and 0x5B, with the same 98.14815% section score. They are the function +extent words: spelling remains 0x1A94 versus target 0x1A90; candidate extent +changes from 0x254C to 0x2594 versus target 0x2574. This is changed unmatched +metadata, not a claim that all data bytes are unchanged. + +Pool check: identical (no narrow strings). Advisory branch mapping: all 351 +pairs preserve their destination and all 71 helper calls correspond. A +register-value-aware linear audit pairs 621 observable store/call/branch +effects. The seven missing byte-boundary differences are resolved; its +remaining three differences are the unchanged count-limit reloads, not a +formal proof or 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 remote, linking, shared-header, assembly, +volatile, register-forcing, dummy-storage or undefined-value changes.