|
|
|
|
@@ -0,0 +1,73 @@
|
|
|
|
|
# data-d5 data lane
|
|
|
|
|
|
|
|
|
|
Baseline 2026-10-02: HEAD and origin/main eecf16f0. All four owned sources equal origin/main before changes. Fresh object build and objdiff report saved in /tmp/data-d5-before.json; instruction-exact baseline in /tmp/data-d5-exact-before.json.
|
|
|
|
|
|
|
|
|
|
| Unit | matched data | total data | instruction exact |
|
|
|
|
|
| --- | ---: | ---: | ---: |
|
|
|
|
|
| SDChannelSelect | 224 | 2960 | 117/129 |
|
|
|
|
|
| ChannelSelect | 192 | 2336 | 101/102 |
|
|
|
|
|
| GCWindow | 248 | 2032 | 47/48 |
|
|
|
|
|
| SDMemory | 200 | 3344 | 63/66 |
|
|
|
|
|
|
|
|
|
|
Pool-first audit: SDChannelSelect 101/101, GCWindow 50/50 and SDMemory 90/90 strings identical. ChannelSelect first divergence index 91, offset 0x653, duplicates WSD_SELECT after its earlier ordinary literal at 0x524. Relocation audit found SDMemory title-pane table at .data+0xc targets A,B,C,D,B_BtnA; source instead uses A,B,B_BtnA,C,D. Correct the typed table initializer.
|
|
|
|
|
|
|
|
|
|
All raw shared .data bytes already equal for SDChannelSelect, GCWindow and SDMemory. Their unequal section scores also involve coarse extracted symbol boundaries and deduplicated weak data. A raw-byte comparison alone does not prove relocations match. No symbols or splits have been changed.
|
|
|
|
|
|
|
|
|
|
## SDMemory
|
|
|
|
|
|
|
|
|
|
Fetched origin before starting; no source changes on origin/main. Attempt 1: corrected the real five-entry title-pane table to A,B,C,D,B_BtnA. Target .data+0x14/+0x18/+0x1c relocations name the corresponding C,D,B_BtnA .sdata literals. Source and target now have identical literal bytes and table targets. Quick gate: data 200 -> 3344/3344, every data section 100%; instruction-exact 63/66 unchanged, objdiff-exact 64/66 unchanged, regressions 0, GATE PASS. No additional data objects or ownership edits needed.
|
|
|
|
|
|
|
|
|
|
## ChannelSelect
|
|
|
|
|
|
|
|
|
|
Fetched origin before starting; no source changes on origin/main. Attempt 1: moved the real scSE_WSD_SELECT definition before preparePageScrolling and referenced that object instead of producing another identical literal. Pool 99/98 -> 98/98 strings, identical; .data raw shared bytes now identical. .data objdiff 32.810795 -> 95.60976; instruction-exact 101/102 unchanged. Remaining semantic reloc differences are exclusively weak interpolation/gui vtables, not strings, arrays or live ChannelSelect/event-handler vtable entries. Source .data 2140 versus target 2144 is trailing linker alignment; source .sdata 73 versus target 80 and .sbss 4 versus 8 also contain only trailing zero alignment. No artificial objects added to fill these gaps.
|
|
|
|
|
|
|
|
|
|
Full clean gate over all four units: GATE PASS, full build ok, DOL SHA1 26116613f624061ba99c8d1a299aaa6efa85670d, regressions 0, forbidden patterns 0, readability problems 0. SDMemory all non-text sections 100% and matched_data 3344/3344. Instruction-exact counts remain 117,101,47,63. Source review is three changed lines plus moving one existing definition.
|
|
|
|
|
|
|
|
|
|
## GCWindow
|
|
|
|
|
|
|
|
|
|
Fetched origin before starting; no source changes on origin/main. Target nm shows LinearIntp<VEC3> destructor and vtable undefined, while the source instantiated a weak LinearIntp vtable at .data+0x650 and Interporation vtable at +0x660, displacing the real MemCardEventHandler vtable from +0x650 to +0x670. These are explicit-specialization ownership differences rather than missing arrays. The existing LinearIntp constructor is instruction-exact, 4/4 instructions, diffs 0.
|
|
|
|
|
|
|
|
|
|
Attempt 1: declare the concrete LinearIntp<VEC3> specialization with an out-of-line destructor, preserving Interporation<VEC3> inheritance. First compilation exposed earlier include order, fixed by including math types under the same GCWindow-only macro. Real MemCardEventHandler vtable returned to +0x650; .data 16.997168 -> 99.55357. The remaining weak Interporation vtable moved to the end. All 47 baseline exact functions preserved.
|
|
|
|
|
|
|
|
|
|
Attempt 2: declare the corresponding concrete Interporation<VEC3> specialization with an out-of-line destructor, keeping its existing fields, initialization and accessors. Both interpolation vtables now have external ownership as the target's LinearIntp relocation proves; no fake zero tables, section attributes or forced references. .data 100%, data 248 -> 2032/2032; all other data sections remain 100%, instruction-exact 47/48 unchanged. Semantic relocations in .data, .rodata and .sdata now agree fully. Every shared-header change is under IPL_GC_WINDOW_CPP, defined only by the owned GCWindow.cpp.
|
|
|
|
|
|
|
|
|
|
## SDChannelSelect
|
|
|
|
|
|
|
|
|
|
Fetched origin before starting; no source changes on origin/main. Pool 101/101 identical. All raw shared bytes of .data, .rodata, .sdata and .sdata2 identical. All non-text semantic relocations before .data+0x960 identical, including every pane-name array, dialog page and compiler switch table. Live button/event/scene vtables start at target 0x980/0x99c/0x9b4 versus source 0x960/0x97c/0x994. Target .data+0x960..0x980 is all zero and has no relocation or symbol. Target trailing deduplicated GUI vtables have zero bytes and no relocations, while source emits real weak PaneManager/EventHandler/Interface vtables. No artificial gap object is permitted.
|
|
|
|
|
|
|
|
|
|
Attempt 1: HermiteIntp<VEC3> specialization inherited the existing Interporation<VEC3> and defined its destructor inline. Preserved all 117 exact functions but generated the two weak vtables at the end, not in the target's earlier anonymous 32-byte gap. .data stayed 35.06984, matched_data 224. Reverted.
|
|
|
|
|
|
|
|
|
|
Attempt 2: use the primary HermiteIntp template with a VEC3 member-init specialization declaration. Preserved 117 exact functions but instantiated init and weak vtables at the end. .data stayed 35.06984, matched_data 224. Reverted because ownership was less faithful to the target's undefined init symbol.
|
|
|
|
|
|
|
|
|
|
Attempt 3: retain concrete HermiteIntp<VEC3> destructor/init out of line but inherit Interporation<VEC3>. Preserved 117 exact functions; only the intermediate weak vtable was emitted at the end. .data stayed 35.06984, matched_data 224. Reverted.
|
|
|
|
|
|
|
|
|
|
Symbol investigation: renamed the two event-handler and scene vtable symbols without changing their extents, using their function-pointer relocations as evidence; .data reached 43.61462, no matched-data gain. Also experimentally split the BaseMask4/Picture_04/ClockPane2 literal-plus-array aggregates at their proven literal lengths and five-/three-pointer table boundaries, preserving section bytes, lengths and all relocations; .data reached 56.610638, no matched-data gain. Reverted all symbols changes. Do not confuse symbol-scoring improvements with exact bytes or acceptance. The anonymous zero gap's exact ownership is not proven, so it remains open rather than being filled with fabricated data.
|
|
|
|
|
|
|
|
|
|
## Data-lane completion audit
|
|
|
|
|
|
|
|
|
|
SDMemory and GCWindow have every non-text section at 100% and matched_data equal total_data. ChannelSelect has identical zero-extended raw non-text section bytes and identical live relocations; remaining .data score 95.60976 comes solely from linker-deduplicated weak interpolation/gui data. Per task step 4, record this extraction limitation instead of deleting correct weak tables or adding zeros. SDChannelSelect has three distinct typed-class attempts plus relocation-backed symbol experiments; its missing anonymous 32-byte zero range and shifted live vtable placement remain unresolved. The code-only non-matches are outside this data-lane assignment and have not been changed.
|
|
|
|
|
|
|
|
|
|
ChannelSelect attempt 2: preserved the Hermite primary template but inherited FrameController directly, with the existing start/end fields instead of the intermediate Interporation base. The intermediate weak vtable disappeared and shifted live event/scene vtables by 16 bytes; .data fell from 95.60976 to 32.810795 despite 101 instruction-exact functions remaining. Reverted.
|
|
|
|
|
|
|
|
|
|
ChannelSelect attempt 3: kept the VEC3 interpolation hierarchy and declared the concrete f32 Hermite specialization with its real fields/init/get and out-of-line destructor, matching the target's external f32 vtable ownership. .data rose to 96.36185, but compiler literal ordering changed .sdata2 to 88.63636 and instruction-exact functions fell 101 -> 100, objdiff code 24736 -> 24272, data 192 -> 104. Rejected and reverted immediately. No failing experiment was committed.
|
|
|
|
|
|
|
|
|
|
All three remaining-data source attempts are recorded for ChannelSelect and SDChannelSelect. GCWindow data is exact after two attempts; SDMemory data is exact after one initializer correction. No code decompilation was undertaken in this lane.
|
|
|
|
|
|
|
|
|
|
## Final clean gate
|
|
|
|
|
|
|
|
|
|
full build: ok
|
|
|
|
|
main.dol sha1: 26116613f624061ba99c8d1a299aaa6efa85670d
|
|
|
|
|
[src/scene/sdChannelSelect/iplSDChannelSelect] objdiff: code 27108/33828 data 224/2960 functions 117/129 fuzzy 99.5051 linked code 0
|
|
|
|
|
[src/scene/sdChannelSelect/iplSDChannelSelect] instruction-exact functions: 117/129
|
|
|
|
|
[src/scene/channelSelect/iplChannelSelect] objdiff: code 24736/25668 data 192/2336 functions 101/102 fuzzy 99.7634 linked code 0
|
|
|
|
|
[src/scene/channelSelect/iplChannelSelect] instruction-exact functions: 101/102
|
|
|
|
|
[src/scene/memoryCard/iplGCWindow] objdiff: code 9172/9964 data 2032/2032 functions 47/48 fuzzy 99.8398 linked code 0
|
|
|
|
|
[src/scene/memoryCard/iplGCWindow] instruction-exact functions: 47/48
|
|
|
|
|
[src/scene/sdChannelMemory/iplSDMemory] objdiff: code 14812/20872 data 3344/3344 functions 64/66 fuzzy 98.9810 linked code 0
|
|
|
|
|
[src/scene/sdChannelMemory/iplSDMemory] instruction-exact functions: 63/66
|
|
|
|
|
regressions vs baseline: 0
|
|
|
|
|
forbidden patterns added (net, per file): 0
|
|
|
|
|
readability warnings (net, per file; must be 0 in the final result): 0
|
|
|
|
|
GATE PASS
|
|
|
|
|
|
|
|
|
|
Fresh comparison confirms no baseline instruction-exact function disappeared. All four string pools identical. GCWindow and SDMemory semantic non-text relocations identical; GCWindow and SDMemory data are 100%. Extab/extabindex, .bss and .sbss are absent except ChannelSelect .sbss, which reports 100%. Config/splits/symbols unchanged in the final diff. Final clean gate log: /tmp/data-d5-final-gate.log. Symbol/relocation evidence: /tmp/data-d5-symbol-relocation-evidence.txt; final semantic relocation comparison: /tmp/data-d5-final-relocs.txt.
|