Compensation-based REF Revert

Overview

Compensation-based REF revert is a state-recovery strategy employed by the Replay lightweight-debugging mechanism in DiffTest-H, a hardware-accelerated processor verification framework. It enables the software reference model (REF) to be efficiently restored to the most recent checkpoint when a fused verification event is found to be buggy, so that the unfused events around the failure point can be reprocessed for debugging.

Motivation

In DiffTest-H's co-simulation, verification events generated by the design under test (DUT) may be fused and buffered for transmission efficiency. Because mismatches between DUT and REF can be detected at any check, the Replay mechanism must revert the REF to the latest checkpoint before reprocessing the relevant unfused events. Naively snapshotting the entire REF at every checkpoint would be prohibitively expensive, particularly in memory usage, since checkpoints occur frequently along the verification timeline.

Strategy

Instead of saving complete REF snapshots, the compensation-based REF revert adopts an incremental modification log:

1. Between any two consecutive checkpoints, only the modifications applied to the REF (e.g., memory writes) are recorded, along with the original pre-modification values.
2. When a fused event is detected as buggy, the REF is restored by rolling back these logged changes.
3. Restoration is performed in reverse order, so the REF state is reconstructed to match the earlier checkpoint without having to persist a full snapshot.

Worked Example: Memory Updates

For memory modifications between two checkpoints, the original memory values are written to the compensation log. To revert:

• Read the log entries in reverse chronological order.
• Write the logged original values back to their respective memory locations.
• After the reverse pass completes, the REF's memory state matches that of the target checkpoint.

This achieves what the source describes as lightweight state recovery, since the per-checkpoint cost is proportional only to the number of modifications, not to the total size of REF state.

Role in the Replay Workflow

The compensation-based REF revert is one of two key challenges addressed by Replay (the other being range determination of unfused events to retransmit). The overall workflow is:

1. Verification events are buffered on the hardware side during fusion.
2. A fused event is reported to the software checker and found to mismatch.
3. The REF's state is reverted using the compensation-based strategy.
4. The hardware is notified to retransmit the relevant buffered unfused events.
5. The checker reprocesses those events to restore instruction-level behavioral detail and pinpoint the faulty instruction and related microarchitectural component.

Benefits

• Low memory overhead: Only modification deltas are stored, not full REF snapshots.
• Deterministic recovery: Reverting in reverse order guarantees a consistent checkpoint state.
• Enables fine-grained replay: Combined with token-based range determination, only the necessary events are retransmitted and reprocessed, avoiding full DUT re-execution.

Related Concept

This strategy is a building block of the broader Replay mechanism; see [Replay (lightweight debugging by event retransmission)] for the full event-buffering, token management, and retransmission design.