LineJoin-CMO: Safe Clean-Line Invalidation without Draining Unrelated Transient State

Authors:

Abstract
A line-scoped cache-block operation (CBO) can be delayed by transient activity that cannot change its target line. Replacing this global drain is not merely an address-filtering problem: an existing target owner must retain a path to completion, new target work must stop entering when ownership clears, and the single-ported directory must remain exclusive through check and update. This paper presents LineJoin-CMO, an effect-scoped quiescence protocol implemented at register-transfer level (RTL) in the synthesizable CVA6/HPDcache CBO.INVAL path. LineJoin-CMO classifies transient state by its remaining effect on the target cache line, drains already-admitted owners, closes target-state admission at the owner-clear boundary, and acquires the directory only after a stable phase transition. Directed tests and a reduced formal model expose distinct bounded non-completion and late-admission failures when the progress and closure boundaries are removed. Across controlled 32-epoch descriptor-ring workloads with 1, 4, and 32 measured overlap opportunities, the first-CBO-to-final-consumer reduction grows from 1.56% to 5.68% and 22.21%. At 32 opportunities, direct fence blocking falls from 2,869 to 158 cycles; endpoint decomposition shows that 1,715 cycles of downstream contention offset part of the local cache-maintenance service saving. Across 24 matched mappings in FreePDK45 and TSMC28, LineJoin-CMO adds one sequential cell and at most 0.287% mapped standard-cell area. At the TSMC28 slow-slow corner and a 3-ns target, the critical delay increases by 2.11% while retaining 0.445 ns slack. These results establish effect-scoped quiescence as a practical, low-cost optimization for clean-line invalidation under overlapping transient traffic.