Automatic Refinement Checking of Pipelines with Out-of-Order Execution

Srinivasan, Sudarshan K.

doi:10.1109/tc.2010.18

Cited by 4 publications

(1 citation statement)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…30 years of effort has gone into developing formal verification techniques for microprocessor design safety and correctness, and a large class of these techniques are established on refinement-based verification [14], [15]. Our central hypothesis is that refinement-based verification developed for safety-verification (functional correctness) of microprocessor designs, and affiliated techniques such as refinement maps, abstractions and microprocessor-specific invariants, can be collectively used to develop efficient and scalable techniques for Spectre invulnerability verification.…”

Section: A Proposed Approach Overviewmentioning

confidence: 99%

A Refinement-Based Approach to Spectre Invulnerability Verification

2022

Self Cite

View full text Add to dashboard Cite

Several mitigations to thwart Spectre attacks have been proposed. However, design errors or trojans can be exploited to circumvent these mitigations. We have developed a highly-automated formal verification methodology that can detect if modern microprocessor designs are vulnerable to Spectre and its variants or prove otherwise. The methodology comprises of a refinement-based formal Spectre invulnerability property, and refinement maps, abstractions and invariants required to efficiently check the property. The methodology was evaluated on 29 benchmarks based on the RSD, which is an open source RISC-V based out-of-order superscalar processor. The benchmarks incorporated two Spectre defenses including software-controlled speculation and data track. Also incorporated were various implementation errors and hardware Trojans such as internally-triggered, rare-circuit-state and time-bombs that induce Spectre. What was evaluated was if the methodology could efficiently prove that an implementation is not vulnerable to Spectre and provide a counterexample otherwise. The methodology provided accurate classification for all benchmarks. The verification times were very efficient, ranging from 30 seconds to 1,500 seconds.

show abstract

Section: A Proposed Approach Overviewmentioning

confidence: 99%