Overview
Randomized Instruction Stream Generation refers to generating processor test stimuli as randomized instruction sequences or streams. In the cited cross-level processor-verification work, the random generator is described as a re-implementation of an existing test generator and as having already demonstrated strong bug-hunting capability. The same work uses the generator in an endless instruction-stream setting rather than as a sequence of isolated test cases.
Role in processor verification
In the DATE 2022 cross-level verification setup, generated instructions are executed by both an RTL core and an instruction-set simulator (ISS), with results compared to find functional differences. The framework includes an instruction-generation component, a coverage observer, an instruction injector, and a comparator. The coverage observer measures functional coverage from the ISS execution state and can provide hints to guide later test generation.
Static random strategy and coverage behavior
The cited work characterizes the baseline random generator as using a static randomized test strategy that does not change over time. This is a limitation for endless instruction streams because the generator cannot be readjusted between separate runs as it could be for individual test cases.
In the reported comparison, instructions produced by the random test generator created substantial peaks for some combinations of instruction groups while other combinations were almost never executed. The paper gives examples where the Special & System : Special & System combination had a very low count, whereas Other : Other was executed very often. The authors therefore describe visible coverage gaps and say the random test generator result appears to degenerate over time.
Relationship to Coverage-guided Aging
The same work compares the static random generator with a generator enhanced by Coverage-guided Aging. The Coverage-guided Aging version had weaker peaks, executed every instruction-group combination in the reported coverage view, and produced clearly visible execution counts for every group. In that comparison, Coverage-guided Aging provided a more balanced result and no visible gaps, complementing the baseline randomized generator for long-running instruction streams.