Only caching use a problem. Indirect branch prediction can be "flattened" at a cost by checking TLB in context switch, presuming the kernel sets it up nicely per process. The other techniques are fine. As usual, secure code needs to be written with timing attacks in mind and they don't change much there.