Pinned loads: taming speculative loads in secure processors

In security frameworks for speculative execution, an instruction is said to reach its Visibility Point (VP) when it is no longer vulnerable to pipeline squashes. Before a potentially leaky instruction reaches its VP, it has to stallÐunless a defense scheme such as invisible speculation provides protection. Unfortunately, either stalling or protecting the execution of pre-VP instructions typically has a performance cost.

One way to attain low-overhead safe execution is to develop techniques that speed-up the advance of the VP from older to younger instructions. In this paper, we propose one such technique. We find that the progress of the VP for loads is mostly impeded by waiting until no memory consistency violations (MCVs) are possible. Hence, our technique, called Pinned Loads, tries to make loads invulnerable to MCVs as early as possibleÐa process we call pinning the loads in the pipeline. The result is faster VP progress and a reduction in the execution overhead of defense schemes. In this paper, we describe the hardware needed by Pinned Loads, and two possible Pinned Loads designs with different tradeoffs between hardware requirements and performance. Our evaluation shows that Pinned Loads is very effective: extending three popular defense schemes against speculative execution attacks with Pinned Loads reduces their average execution overhead on SPEC17 and on SPLASH2/PARSEC applications by about 50%. For example, on SPEC17, the execution overhead of the three defense schemes decreases from 112.6% to 51.3%, from 35.8% to 15.3%, and from 24.8% to 13.2%.