23800 RAR3-p (Compressed) performance problem

[atom]

That is so cool, I haven't even seen that until now. That's because we also improved the crypto library primitives a few weeks ago. And now they become so fast that they are better than then register heavy optimized kernels. Of course that's only for this particulat mode, because it allocates huge register range for every thread. I need to double check that on other GPUs as well, and then simply drop the optimized kernel (and hashcat will fall back to pure kernel).

In regard to -w 4, we will print a warning or so.