The Java virtual machine executes stack-based bytecode. The intensive use of an operand stack has been identified as a major obstacle for static analysis and it is now common for static analysis tools to manipulate a stackless intermediate representation (IR) of bytecode programs. This paper provides such a bytecode transformation, describes its semantic correctness and evaluates its performance. We provide the semantic foundations for proving that an initial program and its IR behave similarly, in particular with respect to object creation and throwing of exceptions. The correctness of this transformation is proved with respect to a relation on execution traces taking into account that the object allocation order is not preserved by the transformation.