NFSU2 on Apple Silicon — static recompilation of a Win32/x86 game to native ARM64

This repository contains the tools and runtime that turn the original Windows executable of Need for Speed: Underground 2 (2004, 32-bit x86, Direct3D 9) into a native ARM64 macOS binary. The game's x86 machine code is lifted to C ahead of time and compiled with clang. There is no Wine, no emulator, no JIT and no Windows involved at run time.

This repository contains no game code and no game assets. It contains the recompiler and the runtime only. You need your own copy of the game; the C code generated from its executable and the resulting binaries are derived from the game and are not, and must not be, distributed here.

Status

The game boots, reaches the menus, and races are playable with sound.

Area	State
x86-32 → C lifter (integer, x87, MMX, SSE/SSE2, 3DNow!, string ops)	23,278 functions, 132.5 MB of generated C
Win32 layer (CRT startup, heap, TLS, threads, files, registry, window)	working
Direct3D 9 → Metal, shader bytecode (vs/ps 1.x–2.x) → MSL	working
DirectSound → CoreAudio, DirectInput (keyboard, mouse)	working
Windowed, fullscreen, Retina-native and 1080p widescreen	working
Double-clickable .app (./pack.sh)	working
Optional "HD" build: 4× MSAA, SSR, ray-traced reflections	working, see GRAFIK_ARASTIRMA.md

Known limitations

• Exactly one build of SPEED2.EXE is supported (see Requirements). Game addresses are hard-coded in the runtime, and the pipeline relies on a property of that particular file (below).
• Videos are skipped: there is no VP6 decoder.
• Keyboard and mouse only; no gamepad or wheel support. No network play.
• Structured exception handling is not implemented (RtlUnwind / RaiseException are stubs).
• x87 is modelled with 64-bit doubles, not 80-bit extended precision.
• One workaround with an unknown root cause remains: the front end's world-view-projection matrix arrives broken and an identity matrix is injected instead (FIXMAT, see run.sh).
• 400 instructions in rarely used opcodes are still emitted as traps. None are hit in the measured race scene.
• Some visuals are unverified against a Windows reference (crowds / skinned meshes, HUD seams). See "Kalan grafik maddeleri" in PROJECT.md.

How it works

SPEED2.EXE (PE32, x86)
   │  tools/reloc2.py     recover the relocation table  → out/relocs.txt
   │  tools/genall.py     recursive-descent discovery + tools/lift.py
   ▼                      x86 → C, one C function per guest function
gen/all/shard_*.c  (128 files) + fntab.c (guest address → native function)
   │  build.sh            clang -O3, in parallel
   ▼
runtime/  (hand-written)  CPU state, guest memory, Win32, D3D9→Metal, audio
   │  link.sh
   ▼
out/nfsu2_boot            46 MB arm64 Mach-O

• Pointers vs. constants. The hard problem of x86 static recompilation is telling a 4-byte address from a 4-byte number. This executable has its relocation directory entry zeroed, but the relocation data is still physically present in an unnamed section. tools/reloc2.py parses it by hand: 91,106 absolute addresses, all of which point into the image. Function discovery and jump-table recovery are seeded from it.
• Guest memory. The 32-bit guest address space is reserved with mmap; guest→host translation is a single ARM64 addressing mode (ldr w8, [x_base, w_addr, uxtw]).
• Indirect calls go through an O(1) table indexed by guest address.
• Code generated at run time. The game writes small x86 thunks to its stack and calls them. runtime/genexec.c detects these and runs them in a small interpreter, handing calls back to lifted functions.
• D3DX effects. The statically linked D3DX9 effect framework took ~47% of guest time when executed as lifted code; runtime/d3dx_hle.c short-circuits parts of it (D3DX_HLE.md).

Path	What it is
tools/lift.py	x86-32 → C lifter, with lazy EFLAGS (flag liveness)
tools/genall.py	function discovery, parallel generation, sharding
tools/reloc2.py	relocation table recovery
tools/argcount.py	stdcall argument counts for imports
tools/isatest.py, tools/fuzzisa.py	instruction-semantics tests and differential fuzzer
runtime/cpu.h, mem.c, dispatch.c, genexec.c	CPU state, guest memory, dispatch, run-time code
runtime/win32.c, win32_ext.c, fileio.c, res.c	Win32 API layer
runtime/com.c, metal.m, vsemu.c	Direct3D 9 / DirectInput / DirectSound COM layer, Metal renderer, shader translation
runtime/audio.m	CoreAudio output
runtime/overrides.c, d3dx_hle.c	hand-written replacements for specific game functions
runtime/enhance.c	HD build only (-DNFS_ENHANCED=1)
runtime/bundle.m, pack.sh	.app behaviour and packaging
tests/boot.c	program entry point (runs the game from its PE entry point)

Verification

• Instruction semantics: 169 hand-written tests, 439 checks (tools/isatest.py → out/isatest).
• Differential fuzzing against real x86: tools/fuzzisa.py assembles random instruction sequences, runs them as real x86-64 code under Rosetta 2 and as lifted C on ARM64, and compares registers, flags and memory.
• Function-level tests: 15 checks against lifted game functions (out/nfsu2_all).
• Reproducible generation: the generated C is a pure function of the executable and the tools; regenerating gives byte-identical output.

./run.sh test runs the first and third.

Performance

Measured on an M1 Pro at 1280×960, same scene and settings, alternating runs (method, raw numbers and caveats in WINE_KIYAS.md):

	This port	Wine 11.18 (wined3d, Rosetta 2)
Race	119 fps (capped by the 120 Hz display)	10–15 fps
3D menus	119 fps	45 fps
CPU time per frame	~12.6 ms	~105 ms

Requirements

• Apple Silicon Mac, macOS 11 or later. Developed on an M1 Pro, macOS 27.

• Xcode Command Line Tools (clang).

• Python 3 with capstone 5.x and pefile (requirements.txt).

• Rosetta 2, only for tools/fuzzisa.py.

• Your own installation of the game, with this exact executable:

  File	SPEED2.EXE, 4,788,224 bytes
  SHA-256	8f6a9ae2c4cfa01181d2fd755d99c6f4b7103f60273908ea1de1e719dfe2a297

  Other builds have not been tested and will not work without changes.

Build and run

Run everything from the repository root.

export NFSU2_DATA="/path/to/Need for Speed Underground 2"   # folder containing SPEED2.EXE
EXE="$NFSU2_DATA/SPEED2.EXE"

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

# 1) relocation table -> out/relocs.txt
.venv/bin/python tools/reloc2.py "$EXE"

# 2) lift every function to C -> gen/all/   (~20 s)
#    OVERRIDE lists the functions that runtime/overrides.c and
#    runtime/d3dx_hle.c wrap; it must match them or the link fails.
OVERRIDE=60fab0,61e390,61f130,5c7ed0,50d510,65ae35,6623ea,66aa3f,66ace2,6691b1,669296,664a2f,663a42,688e89,663275,662edc,660a52,668e0d,664792,66ac27,65addb,65acf5,66d13a,66d303,66d7e9,66e8a9,66f501,66ca38,65ab24 \
  .venv/bin/python tools/genall.py "$EXE" 128

# 3) import argument table -> gen/all/impargs.c
.venv/bin/python tools/argcount.py "$EXE"

# 4) compile the shards, then build the runtime and link
./build.sh && ./link.sh

./run.sh test        # ISA + function tests
./run.sh             # windowed
./run.sh 1080p       # 1920x1080, widescreen-corrected
./run.sh hd          # HD build
./pack.sh            # out/NFSU2.app

run.sh documents the environment switches; most are diagnostics.

Related work

Static recompilation of x86 is not new, and this is not the first project to do it:

• M-HT/SR statically recompiles DOS and Windows game executables (Septerra Core, Battle Isle 3 and others) at assembly level, for targets including macOS arm64.
• sp00nznet/xboxrecomp lifts original Xbox (x86) games to C with a Python + Capstone pipeline very similar to this one.
• N64: Recompiled and XenonRecomp established the approach for MIPS and PowerPC consoles, and are where the method here comes from.
• McSema, rev.ng, RetDec and BinRec are IR-based binary lifters from the research side.

What this project adds is narrower: recovering a relocation table whose directory entry was zeroed and using it to separate pointers from constants; a differential instruction fuzzer that uses Rosetta 2 as the oracle; a hybrid interpreter for x86 code the game generates at run time; and an end-to-end, playable Win32 + Direct3D 9 → macOS ARM64 + Metal port of a 23,000-function 3D game.

Documentation

The engineering notes are in Turkish, as are the code comments.

• PROJECT.md — chronological engineering log: findings, bugs and root causes, measurements.
• D3DX_HLE.md — high-level emulation of the D3DX effect framework.
• GRAFIK_ARASTIRMA.md — graphics research and the HD build.
• WINE_KIYAS.md — comparison against Wine.
• sunum_ss/rapor/nfsu2-rapor.html — illustrated project report.

How this was built

Development was carried out with an AI coding assistant (Claude Code) under the direction of the project owner. Goals, constraints and priorities, observations from playing the game, Windows reference material and direction decisions came from the project owner; implementation, measurement and debugging were done together with the assistant.

Legal

Copyright (c) 2026 Talha Müderrisoğlu. All rights reserved.

This repository is not open source. You may download, build, run and modify it for your own personal, non-commercial use with your own copy of the game. Publishing, distributing or sharing it — or any modified version, port, fork or build of it — and any commercial use require prior written permission. See LICENSE.

Permission: to publish, share or build on this code, ask first by opening an issue in this repository.

This project is not affiliated with or endorsed by Electronic Arts. Need for Speed and Need for Speed: Underground 2 are trademarks of Electronic Arts Inc. The repository contains none of the game's code or data. Screenshots show the game running through this port and are included for documentation; the game content in them belongs to its owners.