Specfuscator: Evaluating Branch Removal as a Spectre Mitigation
Martin Schwarzl, Claudio Canella, Daniel Gruss, Michael Schwarz
FC
Virtual,
March 1-5
2021
Attacks exploiting speculative execution, known as Spectre attacks, have gained substantial attention in the scientific community and in industry with a broad range of defense techniques proposed. In particular, in-software defenses for commodity systems attempt to leave the program structure as is, but defuse every potential Spectre gadget by, e.g., stopping the speculation, or limiting value ranges. While these mitigations disrupt the program flow on every conditional branch, they still contain every single conditional branch instruction. In this paper, we show that one dimension of Spectre mitigations has been overlooked entirely. We explore a novel principled Spectre mitigation that sits at the other end of the scale: the absence of conditional and indirect branches. Our mitigation is based on automatically linearizing the program flow through a special compiler pass, eliminating all conditional and indirect branches. We show that our Spectre mitigation has very clear security guarantees. We explore the feasibility of this unorthodox approach and evaluate its performance in comparison to the more conservative approaches presented so far. We observe that the performance overhead can be low, e.g., 5 %, for certain use cases, being on-par with state-of-the-art mitigations, but very high for other use cases, e.g., and overhead factor of 1000. Our results demonstrate the feasibility of Spectre defenses that eliminate branches and indicate good performance-security trade-offs for Spectre defenses can be achieved by sticking to neither of the extremes.