Answers are grounded in this paper and cite their sources.
Preserved-context processors retain pipeline state across scheduling decisions, but decision-time observations may not suffice to identify a robust action across finite execution horizons. We study primary-eligible initial actions in a RISC-V preserved-context simulator using identical-state counterfactual rollouts, finite-horizon architectural-retirement regret, and restricted observations. We characterize zero robust regret by the existence of an action optimal at every horizon and observation-based robust-decision sufficiency by a common robust-optimal action within each observation class. Across 1280 sampled decision states and horizons 1–48, strict action-preference reversals occur in 141 states under protocol A and 602 under protocol B; minimum robust regret is positive in 125 and 506 states, respectively. Exploratory addition of candidate-wise instruction-fetch (IF) validity eliminates robust-action conflicts in the original run-conditioned classes. However, removing run identity at fixed configuration reveals seven and five conflicts under protocols A and B, respectively. Decision sufficiency on unseen states remains unverified. Under the specified uninterrupted-execution and decision-window conditions, current backend state determines retirement prefixes of two cycles for P4 and three for P5. Controlled hidden-memory witnesses demonstrate that current pipeline observations cannot ensure universal decision sufficiency. Separately, S3 exhibits higher four-cycle post-activation recovery than S0 in 32 of 40 configurations but no improvement in mean full-run architectural retirement. Application-level scheduling benefits are not established.
No takes yet. Share an insight, caveat, or question.
Iuga et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: