I think what we need to do is to test the NOPL instruction and create a
Linux-specific CPUID bit for it. I'll do that and submit to tip:x86/cpu.
I also suspect that we'd really be much better off just fixing the generic NOP tables for the 5-byte nop. As far as I could tell, from all the numbers that have been posted, absolutely _none_ show that there is any point at all to the 2-instruction 3/2-byte sequence.
So instead of having a magic special ftrace-only thing, why not just do it right, and fix the generic 5-byte nop sequence?