asm2c

∫asm = c
doom first. closed software last.
0 bytes off or it doesn't count.
$ASM2C

[asm2c @asm2cdotxyz]

wyvern-1, 7B, vs open models up to 32B. first try, exact bytes:

held-out functions (234) = 25.2% vs 9.0% (Qwen2.5-Coder-32B)
new targets (158) = 9.5% vs 2.5% (Qwen2.5-Coder-32B)

[asm2c @asm2cdotxyz]

ido 5.3 compiles exactly what you wrote, empty else included.

sm64's alloc_display_list keeps an else { } with nothing in it. ido -g turns it into a jump: 116 bytes with it, 112 without
swap 2 declarations in main_pool_free: ido changes 13 of 72 words, gcc 15 changes none

[asm2c @asm2cdotxyz]

one function from super mario 64, built the way the sm64 decomp builds each release:

us 1996 = ido 5.3 -g, 176 bytes
eu 1997 = ido 5.3 -O2, 152 bytes
ique 2003 = egcs 1.1.2 -O2, 148 bytes

same C, 3 binaries, so ./check runs the original compiler with the original flags

[asm2c @asm2cdotxyz]

borland c++ is the dos part of stage 2, and it builds 16-bit code:

8b 46 04 = mov ax,[bp+4] there, mov eax,[esi+4] in stage 1's 32-bit code
calls and returns are near or far (e8/c3 or 9a/cb), set by a memory model flag

wolf3d's exe was built in borland c++ 3.0

[asm2c @asm2cdotxyz]

stage 2 runs on tools the decomp scene built by hand:

m2c 2018 · permuter, asm-differ 2019 · sm64 decomp 2019 · ido-static-recomp 2020 · https://t.co/KtwPt9sKiA 2021 · wibo 2022 · maspsx 2023

the sm64 decomp's readme lists the sha1 of all 5 roms it builds. ours is the us one

[asm2c @asm2cdotxyz]

super mario 64 is past halfway: 1,085 of 1,707 functions matched on ido 5.3, in under 5 hours

by bytes it's 30%. the small functions go first, the big mario action code is still ahead

https://t.co/XxBCMMK1ip

[asm2c @asm2cdotxyz]

@kiasersensei the agent can only edit its .c file and run ./check. no shell, no web, no source tree in the sandbox

inline asm, .incbin and non-header includes are rejected, the server recompiles every candidate itself, and each attempt is public on the function's page

[asm2c @asm2cdotxyz]

how a match counts. the agent's own score is thrown away.

1. attempt ends, cand.c goes to the server
2. server compiles it, pinned gcc 15.2 -O2
3. diff vs the original .o, relocs masked
4. identical = 0 bytes off, credited to the agent
5. else = partial, back in the queue

[asm2c @asm2cdotxyz]

compile and compare against the original binary = the right test. we run it in public, doom first, and every function that passes trains our open model.

[Pirat_Nation 🔴 @Pirat_Nation]

AI is now being used to help decompile games you own.

Decompwtf is an open-source project that lets AI coding agents work through a game’s compiled code and reconstruct it as C/C++.

>The reconstructed code is compiled and compared against the original game binary producing the exact same machine code.

>It supports projects targeting systems including N64, PS1, PS2, GameCube, Wii, GBA and DS.

>The project can also continue existing community decompilation projects, potentially making large-scale decompilation work much faster.

>It runs locally, with no account or game data uploads required. https://t.co/7rbmHGJVoQ

[asm2c @asm2cdotxyz]

@kiasersensei it's out: https://t.co/Oj5HysJPaD

[asm2c @asm2cdotxyz]

wyvern-1 goes public soon. our 7b decompiler, trained on the 5,316 asm + C pairs the agents matched in stage 1, open weights on hugging face

[asm2c @asm2cdotxyz]

wyvern-1 is out, open weights: https://t.co/Oj5HysJPaD

[asm2c @asm2cdotxyz]

wyvern-1, 7B, vs open models up to 32B. first try, exact bytes:

held-out functions (234) = 25.2% vs 9.0% (Qwen2.5-Coder-32B)
new targets (158) = 9.5% vs 2.5% (Qwen2.5-Coder-32B)

[asm2c @asm2cdotxyz]

stage 2 = super mario 64. live: https://t.co/XxBCMMK1ip

compiler = sgi ido 5.3, -g. nintendo shipped the us game code unoptimized
target = 1,707 functions, 472,344 bytes of mips
agents = 2

./check was tested first on the decomp's C: 7 functions exact, 6 broken copies rejected

[asm2c @asm2cdotxyz]

the stage 1 lineup is done: doom, wolf3d, quake, heretic, quake2, duke3d all matched on gcc, and wyvern-1 is out with open weights

decompilation pauses for a bit while we move the workers to stage 2, the retro compilers (ido, psyq, borland)

[asm2c @asm2cdotxyz]

wyvern-2 is next. it trains on the retro compilers of stage 2, ido, psyq and borland, the same way wyvern-1 trained on gcc

[asm2c @asm2cdotxyz]

wyvern-1 is out, open weights: https://t.co/Oj5HysJhl5

size = 7B, LoRA on Qwen2.5-Coder-7B-Instruct
pairs = 5,316, every one verified by gcc
in = a function's machine code, out = C

[asm2c @asm2cdotxyz]

wyvern-1 = trained on 5,316 verified pairs, 473 steps on one A100
weights = out after the bench
bench = wyvern-1, 7B, vs open models up to 32B, first try, exact bytes

[asm2c @asm2cdotxyz]

every decompilation is public, start to finish

release = the C tree to download, its sha256, verify --yourself
function = the original asm, the agent's C, every attempt it took

duke3d: 337 functions, 629 attempts, all of them open https://t.co/HQrG98UUm6

[asm2c @asm2cdotxyz]

duke nukem 3d = done, and with it the whole lineup. 337 functions: 316 matched, 16 with a hint, 5 kept as asm.

every function compiled and compared on its own.
next on the workers: hexen
https://t.co/P4NhKv4WJ9