On in-order processors, without dynamic instruction scheduling, program running-times may be significantly reduced by compile-time instruction scheduling. We present here the first effective certified scheduling compiler that operates over superblocks (it may move instructions across branches), along with its performance evaluation. It is integrated within the CompCert C compiler, providing a complete machine-checked proof of semantic preservation from C to assembly.
Our optimizer composes several passes designed by translation validation: program transformations are proposed by untrusted oracles, which are then validated by certified and scalable checkers. Our main checker is an architecture-independent simulation-test over superblocks modulo register liveness, which relies on hash-consed symbolic execution.
