Overview
RTL simulation executes a processor design at the register-transfer level. In the provided evidence, it appears primarily as the slower but design-level execution target in processor fuzzing workflows. ProcessorFuzz contrasts RTL simulation with ISA simulation: the paper states that ISA simulators are generally much faster for executing a given program than running that program on a processor using RTL simulation, and reports that the RISC-V Spike ISA simulator was, on average, 79× faster than RTL simulation of the RISC-V BOOM processor.
Use in processor fuzzing
ProcessorFuzz uses RTL simulation selectively rather than as the first-stage feedback source. Its design relies on ISA simulation to decide whether a generated test input is interesting, unlike prior work that relied on RTL simulation for that feedback. ProcessorFuzz monitors control-and-status-register (CSR) values across consecutive instructions in the ISA trace; if a CSR value changes, the resulting CSR transition is used to determine whether the input is interesting.
The evidence notes that CSR transitions can be extracted from either an ISA simulator or an RTL simulation of the processor design. ProcessorFuzz chooses the ISA simulator path because it is faster and easier to instrument: the paper says an ISA simulator can be extended using already available trace logic for selected CSRs, while instrumenting RTL designs for coverage tracking requires additional design-specific instrumentation.
Selective RTL execution
In the reported ProcessorFuzz workflow, generated inputs that do not produce a new finite-state-machine state are eliminated using fast ISA simulation before expensive RTL simulation is run. The paper states that ProcessorFuzz launched RTL simulation only for interesting inputs and discarded other generated inputs. It also reports that ProcessorFuzz classified only 33% of generated test inputs as interesting, used the least number of test inputs for RTL simulation, and nevertheless exposed bugs faster in the discussed evaluation.
Practical implication
For fuzzing, the provided evidence frames RTL simulation as valuable but expensive. A common optimization is therefore to use faster ISA-level execution to filter or prioritize inputs, then spend RTL simulation time on inputs expected to explore new processor states.