Philpax

https://fosdem.org/2026/schedule/event/ETLVUA-the_joys_and_horrors_of_nes_dynamic_recompilation/

• Emulator vs recompiler
  • Pros:
    • Simple and flexible
    • Can avoid pitfalls
    • Very easy to debug
  • Cons: Very slow to interpret
• JIT/Dynamic recompiler: the "easy" way
  • When you want to run an instruction, you recompile a block of instructions
  • Pros
    • Usually faster than interpreters
    • Can perform hot-path optimisations
    • Can reuse code blocks
  • Cons
    • Hard to debug
    • Overhead from new code
• AoT/Static Recompilation
  • Recompile all of the code before running it
  • Pros
    • Fastest of all methods
    • Can be recompiled just once, and then reused as required
    • Can be easier to debug
  • Cons
    • Very, very difficul to implement correctly
    • Cannot perform hot-path optimisations
• The project:
  • NES dynamic recompiler
  • C++20, LLVM ORC JIT, SDL
• Small hobby project, so:
  • no self-modifying code
  • doesn't handle all instructions
  • can only rnu a subset of NES games
  • uses an existing interpreter for PPU
  • no audio
• It can run Pac-Man!
• How does it work?
  • For every code block, recompile every instruction into a LLVM function
  • Can restrict the amount to recompile
  • Pass CPU state around as a struct to each code block function
  • Regular instructions are transliterated into LLVM IR
  • For conditional branching instructions, write an early return if possible to avoid recompiling the rest
  • Jump/call writes address of recompiled target
• Horrors:
  • Need to check for CPU interrupts in cycles, which can require cache invalidation