Currently the spec allows implementations to define any cache within the set of standard events. That gets awkward, where one implementation may have an L1 while another has an L1D. I'd suggest the following instead:

caches: L1, L2, LL
TLBs: L1, LL

Implementations would be free to add additional caches/TLBs, but those would be custom events. In both cases, LL can be an alias for L1, L2, or some other cache.

We should define whether LL means the last-level that the core can count, or the last-level in the system, even if the core can't count events for it because it resides outside the core. I lean towards the latter I think?