We describe the first automated approach towards RAM-model secure computation. Towards this end, we define an intermediate representation called SCVM, and a corresponding type system for secure RAM-model computation. Leveraging compiletime optimizations, our approach achieves order-of-magnitude speedup in comparison with naive implementations of RAM-model secure computation. Our approach fundamentally addresses the inherent limitations of circuit-model secure computation considered in almost all previous works. For programs containing dynamic memory accesses whose addresses dependent on secret inputs, existing circuit-model compilers must copy the entire data into the circuit for each dynamic memory access, which is prohibitive when the dataset is large. Our evaluation results show an order-of-magnitude speedup in comparison with the circuit model for reasonably large datasets — and the speedup will be even greater for larger datasets.