Building high performance virtual machines for dynamic languages usually requires significant development effort. They may require an interpreter and one or more compilation phases to generate efficient code. In addition, they may require several static analyses using custom intermediate representation(s). This paper presents techniques used to implement virtual machines for dynamic languages with relatively low development effort and good performance. These techniques allow compiling directly from the abstract syntax tree to target machine code while still enabling useful optimizations and without using any intermediate representation. We have used these techniques to implement a JIT compiler for Scheme. We show that performance of the generated code competes with the code generated by mature Scheme implementations.