I never hid the fact that I iterated with AI. It's pretty hard to commit to an approach if you don't even know if it will succeed. And boy oh boy this was a shot in the dark.
I also respect Wine and OpenBSD and their rules to not have any LLM code in the repo/ports. So my approach was to iterate over solutions, find approach that works, ask AI to "just see if my idea even works", learn what/how/why of that approach, generate some supporting manuals for myself (for example how to decode %gs instructions and how to reencode them) and then write my own code based just on decoding tables (and not LLM code).
I did not know how to work with REX/SIB/ModRM, how to detect prefixes, how to parse every single instruction that contains %gs, how offsets work, how to encode jmp rels, why I want to use JMP *0(%RIP) instead of JMP 32rel on return, why I want to leave stub on SIGUSR (or any other signal). But I do know now.
I was incredibly lucky that patching just these %gs instructions were enough to launch Factorio and play, with broken sound, but anyways! Holy molly this was and is so much FUN!
The approach still feels like cheating, and if it feels the same for you, I perfectly understand.
