Xenosaga Episode I

This project is developed with AI assistance.

A matching decompilation of Xenosaga Episode I: Der Wille zur Macht for the PlayStation 2.

The goal is readable C source that reproduces the original machine code. The Sony SDK and public libraries do not need to be recovered.

Progress

Published source coverage: 7,380 of 7,645 in-scope game functions (96.534%).

Version	Target	Functions	Progress
NTSC-U	SLUS_204.69	3,671 / 3,671	100.000%
NTSC-U	OV01.OVL	879 / 1,081	81.314%
NTSC-U	OV02.OVL	110 / 110	100.000%
NTSC-U	OV10.OVL	298 / 361	82.548%
NTSC-U	OV11.OVL	139 / 139	100.000%
NTSC-U	OV12.OVL	1,788 / 1,788	100.000%
NTSC-U	SSD.IRX	442 / 442	100.000%
NTSC-U	RSSD.IRX	53 / 53	100.000%

Building

The historical compilers are 32-bit Linux executables. On 64-bit Ubuntu, install libc6-i386 and python3-venv. On Windows, use WSL and keep the checkout and build on a Linux filesystem: the old compilers cannot read some Windows filesystem metadata and report Value too large for defined data type.

You need three things the repository does not contain: your own copy of the game, the original toolchain, and a Python environment for the disassembler.

1. The originals

Extract SLUS_204.69, OV01.OVL, OV02.OVL, OV10.OVL, OV11.OVL, OV12.OVL and the IOP/ directory from your own disc image and put them in one folder. config/originals.json names the six files with their sizes and SHA-256; the build refuses anything else, so it tells you immediately if you have a different release. This project reproduces the NTSC-U release.

2. The toolchain

config/toolchain-identity.json: for every compiler and assembler it gives the version, the SHA-256 of each payload, and the provenance needed to obtain or rebuild it. Two of the tools are patched, and both patches are recorded there in full:

Tool	What the build uses it for	Where it comes from
ee-gcc2.96-realconv-lp7	cc1 for the game code (Monolith) and the C runtime (newlib/libgcc)	the published ee-gcc2.96 archive, with twenty-nine documented patches to cc1 (six single bytes that restore the original decimal-literal rounding, two that restore the R5900 short-loop padding, five multi-byte alias-analysis patches, fsv1, fsv3 and fsv4 (%lo-addressed global scalars and struct-field accesses may be reordered in either direction, a non-small-data global's memory reference gets its type's alias set, and gcse load-PRE keeps a fixed scalar load across an in-struct store through a varying pointer), the two fsv5 (qisi) patches (a char sign-extended to 64 bits is lb plus a separate in-place extendsidi2, never one extendqidi2_insn), the three fsv6 patches (symbolic sym(reg) addresses are always legitimate and an address-taken variable gets alias set 0, instead of depending on host heap garbage), the four fsv7 patches (gcse load-PRE: a struct store through a pointer loaded from global S kills a later load of S), the two fsv8 patches (a small-data global with a non-zero initializer gets its type's alias set, as a non-small-data global does), the fsv9 patch (gcse load-PRE: a struct store through a pointer loaded from any global kills a later load of a global), the two fsv10 patches (gcse: a 32-bit union member store through a struct pointer is tested with rtx_varies_p), and the two fsv12 patches (after reload, an address load into a callee-saved register that follows a call in its block is split into lui/addiu))
ee-gcc2.9-991111	cc1 for the SCE SDK scaffold objects only (out of recovery scope, INCLUDE_ASM of the original assembly.	the published ee-gcc2.9-991111 archive, unmodified
ee-as-la29-vsqrt	assembler for every game and runtime object	GNU as 2.9-ee-991111 built from the pinned ps2-ee-toolchain commit with the recorded four-change diff (an overlap-safe memmove, the vsqrt opcode's bits 21-22, labels left before the hazard nops append_insn inserts, and la of a base register plus a 16-bit constant as one addiu)
ee-as-2.9-plain	assembler for the SCE SDK objects	the same commit, unpatched
ps2dev-binutils	linker and objcopy of every image, and modern GAS for data objects and retail blobs	ps2dev binutils 2.45.1, the linker, objcopy and modern GAS

Lay them out under one directory as config/toolchain-identity.json's tools.<id>.dir / .path say. Then:

python3 -B tools/tu/toolchain.py --check

It prints where it looked and what it found, and refuses with the tool's name, both hashes and the recipe when something is missing or different.

3. The disassembler

python3 -m venv .venv && .venv/bin/pip install -r requirements.txt

4. Tell the build where those three are

By environment variable:

export XENO_GAME_DIR=/path/to/extracted-iso
export XENO_TOOLCHAIN_DIR=/path/to/toolchains
export XENO_SPLAT_PYTHON=$PWD/.venv/bin/python

or once, in config/toolchain-local.json (untracked):

{
  "game_dir": "/path/to/extracted-iso",
  "toolchain_root": "/path/to/toolchains",
  "splat_python": "/path/to/.venv/bin/python"
}

Three of the tools are built from source rather than downloaded, so their bytes depend on the host compiler. A build of the same source that assembles the same bytes still reproduces the images: accept it one tool at a time with XENO_TOOLCHAIN_ALLOW=ee-as-la29-vsqrt,ee-as-2.9-plain,ps2dev-binutils. The build then says which check it relaxed.

5. Build

./configure.py      # generate every unit build directory and build.ninja
ninja               # build and compare all six originals
ninja gate          # + per-unit status and the whole-file SHA-256 gate

ninja gate writes build/gate.json. "result": "pass" means every one of the six files came out byte for byte identical to your original.

Public builds use config/objects/data-carves.json to place constants and jump tables emitted by the recovered C functions. Terminal object padding is verified from the published object ranges and original ELF symbols when the private config/tu-build.json is absent. Neither requires private acceptance records.

Contributing

Pick an object from config/objects/<unit>.objects.json that has no source yet, write the C, and build: ninja tells you whether the file still comes out identical. A function counts when its translation unit's source defines it in C and the whole-file gate passes.

Acknowledgments

Thanks to the developers and contributors of the projects on which this work relies:

• splat, spimdisasm, and Rabbitizer for splitting and disassembly.
• m2c for first-pass decompilation of MIPS assembly into C.
• decomp-permuter for the bounded mutation search over a compiled non-exact base.
• ps2dev/binutils-gdb and SSXModding/ps2-ee-toolchain for EE toolchains.
• decomp.me's compiler collection for historical compiler distributions.
• PCSX2 for its documented EE implementation and behavioral reference.

Disclaimer

A legally obtained copy of the game is required. No original game binaries or copyrighted game assets are distributed in this repository. All game rights belong to their respective owners; this project is not affiliated with or endorsed by them.