How are you handling register allocation granularity across architectures? Small changes in regs per thread can shift the limiter in non-obvious ways