Fire Emblem: Seima no Kouseki (聖魔の光石)

A decompilation / disassembly of Fire Emblem: The Sacred Stones — Japanese version (Fire Emblem - Seima no Kouseki, game code BE8J).

It builds the following ROM:

• fireemblem8.gba sha1: 7da0456035366aa18414faa79d8fe7649f03c1ed

This is the JP counterpart of the (essentially complete) US decomp laqieer/fireemblem8u. The two ROMs are compiled from the same Intelligent Systems source by the same compiler, so the bulk of the work is re-linking the US C source against the JP ROM's data layout and decompiling the comparatively small set of region-specific code and data (text, fonts, menus, save versioning).

Goal

Build the ROM byte-for-byte from committed source with baserom.gba removed. A real decompilation reproduces the ROM from matching C (src/*.c), descriptive/data assembly (asm/*.s), and extracted, descriptively-named source assets (PNG/.pal graphics, charmap-encoded text, C struct tables, music) — compiled by an asset toolchain. The original ROM, baserom.gba, is only the verification target of make compare; it is never a build input. Delete baserom.gba and make must still emit the byte-identical ROM (this is what fireemblem8u and pokeemerald already do). See docs/decomp-completion-standard.md for the full standard and the honest audit behind the numbers below.

This is not "asm/baserom.s has zero incbins" (a cosmetic count, gameable by relocating incbins into other files). The single number that matters is build self-containment: remove baserom.gba and see if make still builds. Run python3 scripts/check_selfcontained.py for it, and python3 scripts/calcprogress.py for all scorecard axes, including the strict C/PNG data metric and the supplemental source-form data line.

Status — honest scorecard

FE8J is an in-progress decompilation, but it has reached the foundational milestone: ✅ the ROM builds byte-perfect from committed source with baserom.gba removed. mv baserom.gba away && make emits the byte-identical JP ROM (sha1 7da0456…) — the self-contained build PASSES (the blocking CI gate). The six real axes (target: 100% each — a byte-perfect ROM is necessary but not sufficient; a real decomp is also shiftable and editable):

Axis	Today	Meaning
Build self-containment	100% ✅	the ROM builds from committed source with baserom.gba absent — 0 .incbin "baserom.gba" directives. The only ungameable number, and it's met.
Matching-C functions	99.65% (8650 / 8,680) — 30 still asm (D307: +LoadClassReelFontPalette; the rest are verifier-proven agbcc reg-coloring/RTL walls + large JP reconstructs, kept as descriptive asm like fe8u's own NONMATCHING — D308)	functions whose bytes come from compiling src/*.c plus the handful linked from libc/libgcc archives, over the true JP function total (decompiled + still-as-asm). Earlier this read "99.99% (8,527 / 8,528)" because the denominator was the US ROM's function count — but JP has region-different and JP-only functions the US lacks, so that understated the remainder. Fixed to divide by the real JP total, so 100% is reached only when zero functions remain as asm (the badges above pull this honest number). The Code badge is a separate code-byte measure (~96.7% of code bytes are real-source incl. lib; the remaining ~3.3% is byte-heavy asm), legitimately lower than the function-count. 32 functions remain — JP-only / JP-divergent code that cannot be ported verbatim from fe8u. The ROM cartridge header (src/rom_header.s, fe8u parity) is now real source — its Init entry stub decompiled, its RomHeader* cartridge-header DATA fields (Nintendo logo / title / game-code / checksum) correctly excluded from the function census rather than miscounted as functions. A real-source/lib-link sweep (D288/D289) cleared the newlib stdio members (__swsetup/__smakebuf/__sfvwrite/strcmp, same-size reloc-only), the nullsub_3/nullsub_111 empty-fn return idioms (mov pc, lr / bx lr; nop, committed src/*.s), VerifySramFast_Core + SetSramFastFunc (the SRAM IWRAM-copy cluster, -O1 + static-BSS pin), j_ClearOam (thumb→arm veneer), and the entire newlib locale + malloc clusters — setlocale/localeconv/_setlocale_r/_localeconv_r (locale.o) and _malloc_r/malloc_extend_top/_free_r/_malloc_trim_r (mallocr.o + freer.o), lib-linked with their .text/.rodata/.data sections placed at the JP addresses (the "size mismatch" that had hidden these was an artifact of comparing a whole archive member to a single function, and the .data's "diff" was unresolved relocations — byte-identical once linked). With _dtoa_r lib-linked (its mprec Bigint + libgcc soft-float deps were already linked), the entire newlib/libgcc subsystem is real library source. The remainder is 32 game sub_* functions — mostly register-coloring NEARs (the permuter fleet's domain — several proven to be 2-6 byte residuals that resist every deterministic agbcc lever) plus a few JP-only reconstructions. Much of what D275 called a "compiler-version ceiling" turned out to be agbcc thumb-config knobs, not a foreign compiler: a single agbcc with a -mjp-promote CC flag (built by scripts/build_jp_agbcc.sh, applied per-TU via CC1FLAGS like m4a's old_agbcc) gates thumb.h PROMOTE_MODE (preserve sub-word signedness) + PROMOTE_FUNCTION_ARGS (declaration-order arg extension) — default-off it is byte-identical to stock, so the ~8077 normal TUs are unchanged. It unlocks the s8/s16-hold AND arg-extension-order subclasses plus downstream reg-alloc/LICM/scheduling residuals — incl. AddGorgonEggTrap, D275's flagship "impossible" example (see docs/decisions.md D276/D276b/D276c). The rest is JP-divergent reconstruction + levers + permuter + still-unattributed codegen knobs. See docs/frontier.md.
Extracted data	Strict C/PNG: 79.91% (11,138,009 / 13,937,920); source-form data: 99.26% (13,835,356 / 13,937,920)	The strict badge/decomp.dev-compatible counter is deliberately narrow: bytes in src/*.o C/PNG/charmap/typed data plus libc/libgcc data ÷ measured data bytes. It is not gamed by crediting every non-src/ object, so it stays 79.91%. The new supplemental SOURCE-FORM DATA line in scripts/calcprogress.py explains the apparent 20% path-filter gap: banim/data_banim.o (compressing-linker from banim/*_motion.s + graphics/banim PNG/AGBPAL), sound/songs/midi/*.o (mid2agb from .mid), sound/voicegroups/*.o (editable macro .s), asm/fe6sio.o (mgfembp-built FE6 payload), and the small m4a tables are all committed editable source with no .incbin "baserom.gba" fallback. The source-form line deliberately excludes the remaining 102,564 bytes of opaque raw-incbin/.bin frontier. Named .incbin "baserom.gba" would still be residual and would not receive source-form credit; scripts/check_selfcontained.py currently proves there are 0 such directives in build-input source trees.
Named symbols	85.85% (13,176 / 15,348) — capped: ~1,611 banim_/gfx_/snd_ labels are fe8u's own auto-naming convention (un-named in fe8u itself), so this axis cannot structurally reach 100%; reachable-max ≈ 95.8% of nameable, ~545 blocked (JP-divergent no-fingerprint sub_, inline-__asm__-referenced data_) — D307/W10	labels with meaningful names ÷ total labels. The rest are sub_/data_/nullsub_/sheet placeholders. See docs/decisions.md for the naming-lever history.
Shiftability (data pointers)	gate = 0 ✅ COMPLETE (D309) — the user-ratified criterion audit_pointers.py --true-debt --gate = 0 is met. The last stuck .4byte 0x0800260C literal was a coincidental struct UnitDefinition bitfield word (not a pointer); fixed by typing data_0890915C as struct UnitDefinition[4]. ~14,800 relocated. The 5 remaining "CODE-axis literal pools" are not in the gate — they relocate when their Thumb fn is decompiled (a #2 item). Now also validated by the ported fe8u shiftcheck harness (D317/V1): make shiftcheck — build-address audit → --emit-relocs reloc scan → cross-resource offset scan → differential two-shift; 0 HIGH, wired as a blocking CI step; 8 real cross-resource pointers de-pointered (byte-neutral).	fraction of pointers in extracted data expressed as relocatable symbol references vs baked-in absolute 0x08xxxxxx addresses. Raw INCBIN/​.byte data freezes pointers as constants — on any rebuild that shifts a section they dangle and the game crashes. The fix (D296/D297): rewrite each table into a .4byte Sym relocation block — byte-identical (make compare stays OK) yet shiftable. The literal raw-0x08xxxxxx count (2,867) can never reach 0 — it is dominated by 2,567 coincidental constants (graphics pixels, packed unit-stat bitfields where the level byte is 0x08, sine-table values, sound samples) that are NOT pointers and whose "relocation" would corrupt the shifted game. The honest metric (D304, cross-AI reviewed) is confirmed-real + unclassified real-pointer debt, computed by the fe8u reloc oracle (scripts/fe8u_ptr_offsets.py) plus positive-evidence structural classification (FUNC-interior / ASSET-interior / ROM-header / UnitDef-non-redas-field are provably not pointers). Driven 7,639 → 14,374 over the program; the final session cracked the hex-offset slice class (+97), the impure_data newlib _reent self-pointers (+3), and the gFontgrp Shift-JIS sjisNext glyph linked-list (+234). Residual gate = 12: ~9 deep-blob coincidental, 1 LUT, 2 event-script operands (needs event-script-aware de-pointering). Audit with scripts/audit_pointers.py --true-debt --gate.
Asset editability	measured by SOURCE FORM (per docs/asset_forms.md), not the raw-incbin byte count — a .bin moved into graphics/ passes make compare yet isn't editable (D310 user correction). Done in editable fe8u form: 166/166 event scripts → EVENT_* macros, all proc scripts → PROC_*, 237 UnitDef_* → typed struct UnitDefinition[] (155 + 82 chapter tables at D313/UD1), 21 battle backgrounds → .png (FETSATOOL), 122 palettes → .pal/.agbpal (109 bit-15→.agbpal), 74 chapter maps → .mar/.S, 72 songs → editable .mid + all 93 voicegroups (000-092) → .s (S2: the last 7 — vg035/036/076/077/078/086/092 — converted from opaque .bin/__asm__-literal/blob-embedded data; 0 unresolved pointers) (439 PCM samples already .aif), pixel gfx → .png, battle animation → 201 editable banim/*_motion.s (OAM/script/modes macros) + .png/.agbpal via the fe8u compressing-linker pipeline (D313), replacing 804 opaque data/banim/*.bin, 88 unit-icon move (AP) motion tables → editable src/data/unit_icon_move_motion/unit_icon_move_*_motion.s (.2byte-offset frame/anim macros, ported from fe8u; byte-verified region-same), replacing the opaque data/residual/data_<ADDR>.bin INCBIN wrappers (D316/wave4), ~480 misc/bg/chap-title/gmapunit pixel images → .png (D313/G1-G4), menu string pools → src/menu_def.c literals (D315/C1). Plus 842 dead/orphan .bin removed; scripts/audit_bin_forms.py → docs/bin_audit.md is the reproducible .bin-form miss-tracker (D313/U0). Carve waves 2-4 (D319): +101 AnimSprite_* → typed C struct AnimSpriteData[], +17 AP map-anim/trap objects (316 .bin) → ap.inc macros, +30 ANIMSCR → ANIMSCR_* macros, +147 frontier REDA runs → named struct REDA[] (editable + shiftable), +41 OAM sprite/object-list tables → typed C, +3 worldmap/AP-anim TUs → .s, +m4a/sound tables → .s/typed C — all byte-verified region-same. Waves 6-7 (D321, strict goal "no .bin unless it is .bin in fe8u"): the D319 "reference-parity end-state" claim was RETRACTED — wave 6 proved ~981 real fe8u-editable misses still remained (the banim Img/Pal/Tsa are PNG-derived in fe8u; gFontgrp LUTs were decodable). Tracked committed .bin 4901→3088 (waves 1-5)→2183 (wave 6, PR #75)→1981 (wave 7, PRs #76-79: 112 dead song*.bin deleted, 29 event/proc scripts, 35 banim AP-anim/gfx, 50 typed UnitDefinition + 3 event scripts; 3 shiftcheck HIGH fixed)→1919 (wave 8, D322: 45 INCBIN→typed-C + 6 .png + 8 .agbpal + 3 string pools). The regenerated scripts/audit_bin_forms.py heuristic (corrected classifier, wave 8 / D322) reports MISS 318 / FLOOR 1128 / UNCERTAIN 473 — but 642 is a name-class upper bound, not a work-list (it does not verify bit-exact round-trippability); the careful verified floor is ~159 firm (≈92 compressed-no-bit-exact-recompressor + ≈66 JP-opaque + 1 code-axis) + ~67 soft struct-ptr cascade. A verification pass split the 642 into truly-reducible vs heuristic-false-positive floor — done in wave 8 (D322, 4 verifiers → docs/bin_verification_wave8.md): ~190 truly reducible + ~48 at-parity + ~330–400 floor; the strict goal is not yet reached. Waves 9-28 (D327) then drove the reducible .bin frontier to exhaustion: audit MISS 318 → 11 (tracked .bin 1981 → 1637), every wave byte-exact + make shiftcheck 0 HIGH + CI-green — banim OAM/AnimSprite → typed struct AnimSpriteData[], all 17 mixed JP-LZ blobs hybrid-decomposed to per-sheet .png/.map.bin + typed OAM trailing, pixel-gfx LZ → per-sheet sources, structured offset/palette/OAM/SJIS tables → descriptive .short/.byte, rom_header crt0 → annotated .4byte, ObjectType9 region-diff → u16 palette array. Remaining 11 = documented genuine ceilings (D327): 4 voicegroup/voice tables (blocked on the JP sound-sample-region decomp — 213 unnamed sample pointers), + 7 legitimately-.incbin opaque/compressed/region-diff blobs (LZ-no-clean-sink, opaque LUTs, the 28 KB ch9events JP-frontier blob, an 8 B fragment). fe8u-parity binary floor (kept binary in fe8u too): TSA/tilemap .tsa.bin/.map.bin, compressed region-diff gfx, m4a sample data.	data that is game logic/art should be in its editable fe8u source form (image .png, palette .pal/.agbpal, scripts as macros, tables as typed C), NOT an opaque INCBIN of a .bin. make compare is necessary but not sufficient — it passes on a .bin wrapper too.

Progress is tracked on the project board and charted live on the FE Decomp Portal (alongside fe8u / fe6 / fe7j). See docs/strategy.md for the porting methodology and docs/frontier.md for the single source of truth on what remains (the live, de-inflated work list — do not read layout/nofuncmap_*.tsv as the frontier; it is a stale ~10× classification cache).

Building

You need the ARM toolchain (binutils-arm-none-eabi) and, for C decompilation, agbcc installed into tools/agbcc (same as the US decomp).

# Place your own copy of the original ROM here:
#   ./baserom.gba   (sha1 7da0456035366aa18414faa79d8fe7649f03c1ed)
make compare       # builds fireemblem8.gba and verifies the sha1

Success ends with:

fireemblem8.gba: OK

baserom.gba is NOT a build input — every ROM byte comes from committed source (matching C, descriptive asm, extracted assets, committed data/*.bin). It is used only to verify the result. The self-contained build (the achieved end state):

mv baserom.gba /tmp/ && make           # ✅ builds the byte-identical ROM from source ALONE
mv /tmp/baserom.gba . && make compare   # restore ONLY to verify: sha1 -> OK

Check the current self-containment with python3 scripts/check_selfcontained.py.

Layout

Path	Purpose
baserom.gba	Original JP ROM (you provide; gitignored).
asm/*.s	Carved/descriptive assembly + data. 0 .incbin "baserom.gba" (self-containment is 100%); the bulk is descriptive asm + committed data/*.bin references not yet decompiled to C / extracted as named assets — converted to real source as matching-C and extraction progress (see docs/frontier.md).
src/	Decompiled C (ported/adapted from the US decomp).
include/	Headers (ported from the US decomp).
ldscript.txt	ROM layout; decompiled objects are placed ahead of the incbin.
tools/agbcc	The GCC 2.95 ARM compiler (install locally; gitignored).
docs/	Strategy & methodology notes.

Reverse engineering

For the remaining region-different functions (text/font/menu/save) that must be hand-decompiled against the JP ROM, two decompilers are wired up headlessly over MCP so Claude Code (and the autonomous loop) can ask for pseudo-C of a JP address: IDA Pro Hex-Rays (server ida, the workhorse — make ida-db) and Ghidra (server ghidra, an open-source cross-check — make ghidra-db). Both are fed the project's own fireemblem8.elf. To close the last gap when a ported function compiles close but not byte-exact, decomp-permuter (upstream, which now supports ARM32/Thumb + agbcc) is wired up via scripts/permuter/. Setup, rationale and usage are in docs/reverse-engineering.md and docs/decisions.md (D6). The IDA/Ghidra/permuter installs and databases are local/gitignored; only the scripts/config are tracked.

License

Decompilation infrastructure mirrors the US decomp. Game assets/data are the property of Nintendo / Intelligent Systems; no copyrighted ROM data is committed.