Compare commits

...

1 Commits

Author SHA1 Message Date
coah
5120d6977d data(iplKeyboard data 1176/1184) (data-d1b) (worker 9e9e6cc1) 2026-10-02 05:42:11 +00:00
2 changed files with 53 additions and 3 deletions

View File

@@ -219,9 +219,7 @@ namespace ipl {
u8* usedSystemDict = System::getZiDicData(i);
u8* usedOemDict = System::getZiDicData(oemIndex);
if (usedSystemDict == NULL) {
usedSystemDict = systemDict;
}
usedSystemDict = usedSystemDict != NULL ? usedSystemDict : systemDict;
if (usedOemDict == NULL) {
usedOemDict = oemDict;
}

View File

@@ -44,3 +44,55 @@ Data-lane audit: OSUtf has no unresolved section; all functions exact. Controlle
## Final verification
Clean non-quick four-unit gate PASS. Full build ok; DOL SHA1 26116613f624061ba99c8d1a299aaa6efa85670d; regressions 0; forbidden patterns 0; readability warnings 0. Fresh pool_diff checks all identical. Fresh getter ctxdiff 18/18 instructions, diffs 0. Final exact counts OSUtf 4/4, keyboard 31/32, controller 84/84, NHTTP 22/22. Final matched data OSUtf 48704/48704, keyboard 8/1184, controller 80/696, NHTTP 16/720. Final matched code OSUtf 568/568, keyboard 4752/6024, controller 5708/5708, NHTTP 3180/3180. The gate passes but the all-data-100% objective is not met for the latter three units; no claim of completion for those sections. All remaining genuine unmatched functions have three distinct attempts recorded. Final full gate output is /tmp/data-d1.final-gate.txt.
# Round 2, after #730
Fresh leaf HEAD bcb2ac0e equals fetched origin/main at start. Initial quick three-unit gate PASS, DOL SHA1 correct, zero regressions. Keyboard data 8/1184, exact 31/32, code 4752/6024; controller data 80/696, exact 84/84, code 5708/5708; NHTTP data 16/720, exact 22/22, code 3180/3180. All three pool_diff checks identical. Rechecked ELF bytes, nm, objdump symbol tables and relocations. Prior round's byte/relocation findings remain reproducible.
Inspect the installed objdiff 3.4.5 data comparison algorithm to distinguish relocation mismatches from legal linker tail alignment; do not alter comparison settings or tool code to hide mismatches. Investigate source changes and extraction metadata only when supported by byte/relocation evidence.
Round 2 keyboard create attempt 1: replace the second loop's parallel OEM induction variable with the meaningful expression `i + EZTX_LANG_MAX`. Objdiff create 94.50944%, 313/318 instructions, no exact function loss. Compiler strength-reduces both indices into one byte offset and removes five instructions instead of recovering the target's independently maintained OEM index. Not accepted.
Objdiff 3.4.5 diff_data_section compares bytes up to maximum data symbol extent and only uses raw byte match percentage when every original relocation matches. This explains keyboard .data's 25% despite identical bytes: create jump-table relocations fail. The final keyboard data string ends at 1161 on both sides, so those seven section tail bytes do not obstruct .data matching. NHTTP original last coarse blob includes tail alignment and extends to 704 while built last literal ends at 701, so its 99.786476% score is the byte ratio 2*701/(704+701). Controller's extra weak Interface table extends built symbol extent to 752 against original 616. These are specific measured causes, not guessed missing objects. Do not change symbol sizes or suppress the real extra source vtable to fake matching.
Round 2 keyboard create attempt 2: express both OEM indices as `i + EZTX_LANG_MAX`, removing parallel induction variables. 309/318 instructions, create 92.28616%, no exact function loss. Both loops strength-reduce OEM addressing to the system byte offset plus 40, unlike target's separate index addition and shift. Reverted.
Round 2 keyboard create attempt 4, memo_declaration_order: src 0x4f4 base 0x4f8 insns 317/318; create 94.157234%; code 4752, data 8, exact-name functions 31.
Round 2 keyboard create attempt 5, memo_copy_initialization: src 0x4b8 base 0x4f8 insns 302/318; create 87.8522%; code 4752, data 8, exact-name functions 31.
Round 2 keyboard create attempt 6, dictionary_fallback_order: src 0x4f4 base 0x4f8 insns 317/318; create 93.83648%; code 4752, data 8, exact-name functions 31.
Round 2 keyboard create attempt 7, dictionary_ternary_fallback: src 0x4fc base 0x4f8 insns 319/318; create 93.93082%; code 4752, data 8, exact-name functions 31.
Round 2 keyboard create attempt 8, oem_load_after_system_fallback: src 0x4f4 base 0x4f8 insns 317/318; create 94.57861%; code 4752, data 8, exact-name functions 31.
Round 2 keyboard create attempt 9, signed_counter_unsigned_oem: src 0x4f4 base 0x4f8 insns 317/318; create 94.18239%; code 4752, data 8, exact-name functions 31.
Round 2 keyboard create attempt 3: use unsigned loop induction variables for both dictionaries, retaining parallel OEM indices. Output unchanged at 317/318 instructions and 94.18239%; exact functions remain 31. Reverted.
Round 2 controller vtable attempt 1: make the abstract Interface destructor an inline non-pure destructor under a source-only header guard. Abstractness remains enforced by decide and setForceInvalid; this tests whether the pure destructor caused unwanted table emission. All 84 original functions remain 100%, but extra Interface table remains and .data stays 92.03601%. Reverted both source and header; no justification for changing destructor semantics without a match.
Round 2 keyboard create attempt 10, system_ternary_fallback: src 0x4f8 base 0x4f8 insns 318/318; create 94.00944% ; code 4752, data 1176, exact-name functions 31.
Round 2 keyboard create attempt 11, oem_ternary_fallback: src 0x4f8 base 0x4f8 insns 318/318; create 94.00944% ; code 4752, data 1176, exact-name functions 31.
Round 2 keyboard create attempt 12, system_inverted_ternary: src 0x4f4 base 0x4f8 insns 317/318; create 94.35535% ; code 4752, data 8, exact-name functions 31.
Round 2 keyboard create attempt 13, oem_inverted_ternary: src 0x4f4 base 0x4f8 insns 317/318; create 94.35535% ; code 4752, data 8, exact-name functions 31.
Round 2 candidate: retain attempt 10's ordinary system dictionary fallback selection `usedSystemDict != NULL ? usedSystemDict : systemDict`. OEM selection remains unchanged. The resulting create has 318 instructions, and all twelve destination switch-table relocation addends now exactly match the original object. Both .data byte extents are 1161, and every one of the 85 original .data relocation records matches; objdiff .data is 100%, raising matched_data from 8 to 1176 of 1184. No string, symbol, section ownership, source object, or data byte is altered; no padding or assembly is introduced. This is a genuine relocation-data match from a readable conditional expression, even though create remains instruction-nonexact at 94.00944%. Instruction-exact count remains 31/32. Attempts 11 through 13 are not retained. The data-lane task permits improvements with unchanged exact counts; the generic worker count-growth criterion is not met.
Candidate quick three-unit gate PASS, zero regressions, zero forbidden patterns, zero readability warnings, correct DOL SHA1. Keyboard .data 100%; .sdata remains 85.71429%, .sdata2 100%. Controller .data 92.03601%; other data sections 100%. NHTTP .data 99.786476%; .sdata and .sbss 100%. No original extab/extabindex, .bss, or additional small-data sections are present in the owned objects.
Remaining keyboard .sdata contains `arc\0` at offset 0, followed by empty 16-bit wchar string at offset 4. Source literal is 2 bytes, extraction labels it as a 4-byte object including two alignment zeros. All instructions using it still match exactly. Changing L"" to an artificially longer null literal or enlarging its symbol would fabricate storage; neither is accepted. NHTTP and controller residual section differences remain exactly the extraction alignment/weak-vtable findings above. They contain no unmatched functions. All failed experimental source and header changes were restored before the candidate gate.
Attempts 4-9 detail: 4 reverses the two MemoSetting declarations; 5 uses direct initialization for the loaded setting and copied setting; 6 reverses system/OEM fallback statement order; 7 expresses both fallback selections as ternaries; 8 postpones the OEM pointer load until after system fallback selection; 9 moves OEM induction initialization outside the first loop and uses an unsigned system index. Each was independently compiled, measured, and reverted. Attempts 10/12 use respectively non-null/null-first system ternaries; 11/13 use respectively non-null/null-first OEM ternaries. Only 10 is retained. The single remaining unmatched function, keyboard create, has thirteen distinct source-level experiments in this round, in addition to the three original attempts. Controller and NHTTP have zero remaining unmatched original functions.
Final fetch during verification reached f6fd2780; owned three sources do not differ between this branch's base and fetched origin/main. The worktree remains on its assigned leaf branch.
## Round 2 final clean gate
Final non-quick gate over all three units PASS. Full 43U build succeeds and DOL SHA1 is 26116613f624061ba99c8d1a299aaa6efa85670d. Regressions 0; forbidden additions 0; readability warnings 0; all pools identical. Keyboard exact functions 31 -> 31, matched code 4752 -> 4752, matched data 8 -> 1176 of 1184; .data 100%. Controller exact 84 -> 84, code 5708 -> 5708, data 80 -> 80 of 696. NHTTP exact 22 -> 22, code 3180 -> 3180, data 16 -> 16 of 720. No instruction-exact function count dropped. Fresh relocation comparison confirms every original keyboard .data relocation matches. Final gate output /tmp/data-d1-r2.final-gate.txt.
Incomplete data sections: keyboard .sdata 85.71429% from extracted empty-string extent; controller .data 92.03601% from linker-discarded weak Interface vtable and unit-boundary alignment; NHTTP .data 99.786476% from last extraction blob covering three tail alignment zeros. No symbol size or section boundary has been changed to improve the score. The all-data-100% goal remains incomplete. Remaining function create is 94.00944%, 318/318 instructions with 210 instruction differences; register allocation, dictionary-loop instruction scheduling, and MemoSetting temporary stack placement still differ. Every remaining function has at least three distinct logged attempts; all thirteen are retained as evidence, with only the data-exact candidate in the source diff.