Non-Speculative Load-Load Reordering in TSO

Abstract

In Total Store Order memory consistency (TSO), loads can be speculatively reordered to improve performance. If a load-load reordering is seen by other cores, speculative loads must be squashed and re-executed. In architectures with an unordered interconnection network and directory coherence, this has been the established view for decades. We show, for the first time, that it is not necessary to squash and re-execute speculatively reordered loads in TSO when their reordering is seen. Instead, the reordering can be hidden form other cores by the coherence protocol. The implication is that we can irrevocably bind speculative loads. This allows us to commit reordered loads out-of-order without having to wait (for the loads to become non-speculative) or without having to checkpoint committed state (and rollback if needed), just to ensure correctness in the rare case of some core seeing the reordering. We show that by exposing a reordering to the coherence layer and by appropriately modifying a typical directory protocol we can successfully hide load-load reordering without perceptible performance cost and without deadlock. Our solution is cost-effective and increases the performance of out-of-order commit by a sizable margin, compared to the base case where memory operations are not allowed to commit if the consistency model could be violated.

Maximizing Limited Resources: a Limit-Based Study and Taxonomy of Out-of-Order Commit

“…The insight of this study is that selective speculation, preserving or relaxing of one more conditions necessary for out-of-order commit, could lead to more efficient implementations. In fact, recent work [15,24] has shown that this might be the case for load instructions.…”

“…This is especially true if the potential benefits can be tapped with much more efficient and selective mechanisms to guarantee correctness, instead of bulk checkpointing and rollback. Already, such non-speculative mechanisms have been proposed [24].…”

“…Evading out-of-order commit conditions. [24] provides a coherency protocol level solution such that the load→load reordering can be evaded non speculatively under TSO. In their design, it is no longer necessary to squash and re-execute speculatively reordered loads if their reordering can be hidden by the coherency protocol.…”

Non-Speculative Store Coalescing in Total Store Order

“…In general, a deadlock avoidance mechanism cannot simply rely on increasing associativity (e.g., with the addition of a fully associative victim cache) as this only makes the deadlock more unlikely, but not impossible. Figure 8: Directory-eviction deadlock Note, here, that the resource conflict does not concern reads (loads), as the inability to evict means that read will be delayed until the atomic-group write completes-or the read will be performed as uncacheable [17].…”

“…Our work is largely inspired by concepts introduced in the following works [10], [7], [5], [8], [25], [26], [6], [17]. We aim to show that store!store reordering in TSO, to the extent it is imposed by coalescing, can be achieved non speculatively.…”