I am frustrated that we are going around in circles with respect to modulation of CMOs

Current proposal, hoping to break this deadlock: Privilege for CMOs

Interception, Modulation, and Mapping of CMOs

Original Proposal: CSRs ansd system calls.

My (Ag's) original proposal looked something like this:

• CMO instructions that contain CSR operand, along with and address or a (set,way) cache entry number that indicated what to flush
• A CSR operand for each such CMO instruction, that contained an encoding that indicated which caches aned branch predictors need

however, it is necessary to accompany this with a system call:

1. since the user cannot write such a CSR directly
2. since different software systems may allow may allow (some) users to perform a CMO, while the same or other software systems may disallow (some) users from performing that same CMO
   • i.e. the privilege required for a CMO depends on the system software. It is NOT KNOWN to CPU hardware or the ISA
3. since there needs to be a mapping between abstract user level CMO's and the operations that the hardware actually performs

Mapping

local cluster HW coherent MOESI SW coherence between clusters

SW P -> C

MOESI

• flush all dirty data in local cluster to the poc(P,C) MESI
• no thread migration
  • flush local CPU only
• thread - flush all cluster

Point_of_Unification = pocvg(P.I,P.D) * pocvg(P*.I,P*.D)

Point_of_Coherence = pocvg(P1.D,p2.D;address)

Point_of_Persistence = pocvg(P1,NVRAM) or pocvg( *

Point_of_Serialization = per address

• FENCE.COMPLETION = persistence / SW coherency / MMIO