Fast and Accurate Simulation using the LLVM Compiler Framework?

Abstract. Development of future generation computer architectures re-quires fast and accurate simulation tools that allow to test, verify, and analyze the behavior of the given architecture along with the intended workload. We present a simulation framework based on a structural ar-chitecture description language that uses the open source compiler in-frastructure LLVM to dynamically translate instruction sequences of the simulated architecture into machine instructions of the host machine. We show that the optimizations in the simulator and the LLVM com-piler lead to an outstanding runtime performance: A 5-stage MIPS core is simulated at a peak performance of up to 800 MHz.