mirror of
https://github.com/Druthulu/BFM-decomp
synced 2026-09-26 21:36:06 -04:00
Drew T
28c23f4be8
- func_8015F9A4 byte-matched + banked inline in ov_SC01_077 (whole-binary gate d19c9580, BYTE-IDENTICAL); a near-clone of the demo's func_8015F89C fnptr close -> fnptr-call class confirmed reliably closeable (2/2) - firmed the rate: loop reconstructs CORRECT bodies 4/4, but byte-close is gcc-quirk bound. func_8014C308 (relocs=0, x134) structurally perfect, v0/v1 regalloc only (permuter stuck 455->140); func_80130AF0 (x134) body-100% but S10 delay-slot remat; func_8016B4F8 phantom -O2 frame. Drafts in .run/demo/ (gitignored) - new reusable idioms (hand-matching-process.md S2): mask-local (defeats lh->lhu fold), shared-ret0 goto (cross-jump clustering + branch polarity) - harvest sizing (S6): tractable easy classes are LOW-reach (fnptr 0.1%, void-widen 0.3%); the x134 yield is locked in quirk-heavy STRUCTURAL_MISS (7.6%)/PERMUTER_CLASS (3.6%). Permuter route refuted (T6 callee-inlining + func_8014C308 slow). Lever = Ultracode guided-hand-matching wave on the high-reach tractable subset - plan revised (Drew, A-only): harvest tractable first; giant func_80144B9C deferred