mirror of
https://github.com/RetroPortingToolKit/psxrecomp
synced 2026-10-26 02:11:36 -04:00
Matthew Stanley
414314d412
The old delta-sweep matched export-table pointers against 0x27BD prologues, but those pointers point at mid-function DISPATCH entries, not prologues (measured: 0/29 hit a prologue at the true base). It scattered bases by 2-16KB and decoded data as code. Fix: j/jal targets are BASE-INDEPENDENT (0x80000000 | (imm26<<2), since all game code lives in 0x800xxxxx). At the true base, the maximum number of intra-overlay jal targets land on a real prologue. recover_base() votes every (jal-target, prologue-offset) pair for base t-o within the pointer-containment window; the peak is unique and sharp (~2x margin). Records are emitted page-aligned with a FILL prefix, because the runtime keys shards by a page-aligned region_start and compile_overlays takes load_addr verbatim. Weak-signal files are skipped (score<4 or <1.5x runner-up) -- safe coverage loss under the per-function code_crc dispatch guard. Also dedup overlays by content CRC (position-fixed BINs are reused across AREA folders). Validation: generic output is byte-identical (base, region envelope, seed set) to the live-proven tomba1_extract.py hand tool for ALL 22 Tomba 1 overlays, with no hardcoded constants. Tomba 2's 21 A*.BIN overlays converge on a single consistent base 0x80108F9C (region 0x80108000 +3996). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>