Decomposing the load-store queue by function for power reduction and scalability

Abstract

Abstract-Because they are based on large content-addressable memories, load-store queues (LSQs) present implementation challenges in superscalar processors, especially as issue width and number of in-flight instructions are scaled. In this paper, we propose an alternate organization of an LSQ that separates the time-critical forwarding functionality from checking that loads received their correct values. Two main techniques are exploited: 1) the store forwarding logic is only accessed by those loads and stores that are likely to be involved in forwarding, and 2) the checking structure is banked by address.The result of these techniques is that a collection of small, low bandwidth structures can be substituted for the large, high bandwidth structures used in conventional designs. By our calculations, these proposed techniques reduce LSQ dynamic power by a factor of 3-5 while achieving equivalent performance. Furthermore, we explore applying these techniques to processors requiring only statistical correctness guarantees where another factor of two power reduction is achievable.

NoSQ: Store-Load Communication without a Store Queue

“…Proposals can be grouped into three general classes. The first class maintains the age-ordered store queue structure but uses partitioning, filtering, hierarchy, dependence speculation, and speculative forwarding through the primary data cache or other structures to reduce the frequency of associative store queue search or the number of entries examined per search [2,5,12,18,20]. A second class avoids associative search by abandoning the conventional age-ordered structure and replacing it with a cache-like address-indexed structure [6,18,21,24].…”

“…Associative search constrains the scalability of the store queue, which in turn constrains the scalability of the entire instruction window. To address this challenge, recent work has proposed to reduce both search frequency and the number of entries that must be searched [2,5,12,15,18,20], to replace the fully-associative age-indexed store queue with a set-associative address-indexed forwarding structure [6,21,24], or to maintain the age-ordered structure but replace associative search with speculative indexed access [19,22]. This paper presents NoSQ (short for No Store Queue and pronounced like "mosque"), a microarchitecture that implements in-flight store-load communication without a store queue or any other intermediary structure.…”

Slackened Memory Dependence Enforcement: Combining Opportunistic Forwarding with Decoupled Verification

“…This L1 structure is backed up by a much larger second-level (L2) structure to correct/complement the work of the L1 structure. The L1 structure can be allocated according to program order or execution order (within a bank, if banked) for every store [1,8,24] or only allocated to those stores predicted to be involved in forwarding [3,17]. The L2 structure is also used in varying ways due to different focuses.…”

“…The L2 structure is also used in varying ways due to different focuses. It can be banked to save energy per access [3,17]; it can be filtered to reduce access frequency (and thus energy) [1,19]; or it can be simplified in functionality such as removing the forwarding capability [24].…”

“…Besides the difference in focus, the mechanisms we use in the speculation and the recovery from mis-speculation also differ from this prior work. In a two-level disambiguation approach [1,3,8,17,24], the fundamental action is still that of an exact disambiguation: comparing addresses and figuring out age relationship to determine the right producer store to forward from. The speculation is on the scope of the disambiguation: only a subset of the stores are inspected.…”