OpenUT99-Plus
Unreal Engine 1, rebuilt from scratch
Loading...
Searching...
No Matches
0006 — Execute the game's UnrealScript bytecode in a VM

  • Status: Accepted
  • Date: 2026-11-10 (accepted 2026-11-11)

Context

UT's game logic is UnrealScript bytecode inside the retail .u packages (script.md). Measured on System/*.u (469e):

Package group Functions Native functions Bytecode (memory bytes)
Gameplay: BotPack, UnrealShare, UnrealI, multimesh, relics, de, SkeletalChars 5,041 0 776,555
UI: UWindow, UMenu, UTMenu, UBrowser, UTBrowser, UTServerAdmin 2,251 0 524,599
Engine core: Core, Engine 1,446 308 145,147
Other: IpDrv, IpServer, UWeb, udemo, Editor, OldUnreal469c 418 75 86,733

241 distinct native function indices are declared (highest 3971). The bytecode decoder already walks all 1.58 million tokens of the install exactly.

Options:

  1. VM: load classes from the user's packages and run their bytecode; implement in C++ only the native functions (≈ 383) and native classes' behaviour (actors, physics, rendering, networking).
  2. Native reimplementation: rewrite the game classes (≈ 9,000 functions) in C++.

Decision

Option 1, a bytecode VM.

  • ADR 0004 forbids translating UnrealScript line by line; option 2 would mean rewriting 1.3 MB of gameplay and UI logic "from behaviour", the largest and least verifiable path.
  • The VM runs the original logic unchanged, which is what network compatibility (ADR 0005) needs: replication follows the classes' own replication blocks and state.
  • Mods and the 120 cached third-party packages are script too; only a VM runs them.
  • The work concentrates on a small, testable surface: ≈ 383 native functions, specified one by one from their declarations, the script that calls them, and observed behaviour.

Consequences

  • New engine modules: class/property layout (from the property declarations), object instances with default properties, the VM, native function table.
  • The VM is verified against the original by observable behaviour (game state, logs, network traffic), not by reading native binaries.
  • Performance: the original ran this bytecode on 1999 hardware; an interpreter is acceptable.