Repository navigation
mkw: Match KartState::init, improve KartCollide::checkNeighborhood, decompile RaceinfoPlayer_getLapSplit - #12
ManuDigitalSun wants to merge 3 commits into
Conversation
Match the 56-byte function at 0x8059455C. Verified with Objdiff at 100.0% and: ninja all_source progress build/RMCP01/report.json AI assistance: Used an AI coding assistant for build-environment troubleshooting, report/candidate analysis, target-assembly interpretation, and drafting the implementation. I reviewed the change and independently verified the final match and full local build.
Enable and refine the existing partial implementation. Objdiff: 77.65% -> 99.60% Function size: 0x224 / 548 bytes (matches target). Original partial implementation by em-eight (7974069). Known remaining mismatch: equivalent float constants use a different read-only-data layout. The target uses @5034 and lbl_1_rodata_2020; the current build uses anonymous local literals at different offsets. AI-assisted: OpenAI Codex was used for assembly analysis, source-shape experiments, and Ninja/Objdiff build iteration.
Objdiff: 98.910255% (312/312 bytes). AI-assisted: OpenAI Codex was used for target-assembly analysis, source-shape experiments, and Ninja/Objdiff build iteration.
Report for RMCP01 (fa12730 - bf6a07a)📉 Matched code: 5.77% (-0.00%, -132 bytes) ✅ 1 new match
🥀 1 broken match
📈 3 improvements in unmatched items
|
| } | ||
| } | ||
|
|
||
| extern "C" void RaceinfoPlayer_getLapSplit(RaceManagerPlayer* player, u32 lap, Time* split) { |
There was a problem hiding this comment.
Why is extern C used here? Why not the C++ name? Maybe the LLM used does not understand code warrior mangling?
There was a problem hiding this comment.
Maybe, it's a good question to understand how the LLM thinks about these problems. I asked the model about it:
Prompt: In the RaceManagercpp RaceInfoPlayer_getlapsplit, why did you use extern C instead of the c++ name? Do you not understand code warrior mangling?
Answer:
I understand CodeWarrior mangling. extern "C" was a linkage workaround, not proof that the original function was written as C.
The target symbol is literally:
RaceinfoPlayer_getLapSplit
without CodeWarrior mangling. A normal C++ definition such as:
void RaceManagerPlayer::getLapSplit(u32 lap, Time* split)
would emit a mangled symbol resembling:
getLapSplit__Q26System17RaceManagerPlayerFUiPQ26System4Time
and would not replace the plain target symbol by itself.
However, the ABI strongly suggests it behaves like a member method:
r3 = RaceManagerPlayer* this
r4 = lap
r5 = Time* split
So the honest answer is: I used extern "C" to make the source emit the target label while we had not recovered the proper C++ declaration/name mapping. It is technically valid for matching the ABI and label, but it is not ideal recovered source.
A cleaner reconstruction would be to add:
void getLapSplit(u32 lap, Time* split);
to RaceManagerPlayer, then determine how this project expects the member’s CodeWarrior-mangled output to be mapped/aliased to the existing target label. We should not blindly change it now, because the emitted mangled symbol would stop covering RaceinfoPlayer_getLapSplit unless we also resolve that symbol-mapping issue
| if ((kartColInfo.flags & (COL_FLAG_WALL_AT_LEFT_CLOSER | COL_FLAG_WALL_AT_RIGHT_CLOSER)) == 0) { | ||
| if (fabsf(hitbox.bsp->pos.x) > 10.0f) { | ||
| if (hitbox.bsp->pos.x > 0.0f) { | ||
| ((vu32&)kartColInfo.flags) |= COL_FLAG_WALL_AT_RIGHT_CLOSER; |
There was a problem hiding this comment.
Since this is not 100% matched anyways I would not include the vu32 casts in the commit
2 participants
Summary
KartState::init(56 bytes @ 0x8059455C) — 100.0% Objdiff match.KartCollide::checkNeighborhood— enables and refines the existing partial implementation by em-eight (7974069), 77.65% → 99.60% (0x224 / 548 bytes, matches target size). Known remaining mismatch: equivalent float constants use a different read-only-data layout (target uses@5034/lbl_1_rodata_2020; current build uses anonymous local literals at different offsets).RaceinfoPlayer_getLapSplit— 98.910255% (312/312 bytes).Verification
Verified with Objdiff and a full local build:
AI assistance disclosure
AI-assisted: OpenAI Codex was used for target-assembly analysis, source-shape experiments, build-environment troubleshooting, and Ninja/Objdiff build iteration. All changes were reviewed and the matches independently verified by me.