Efficient online optimization by utilizing offline analysis and the safeTSA representation

Conventional mobile-code representations, e.g. Java bytecode, provide machine-independence and type-safety, but do so at the expense of performance. This performance hit can be taken in the form of decreased throughput or in increased latency. SafeTSA was designed to reduce this performance hit, especially when producing high-quality optimized machine code. It does this by utilizing SSA form and thus shifting dataflow analysis effort from the online JIT compiler to the offline producer of the safeTSA program. This paper describes the extension of safeTSA to support a greater shifting of optimization effort away from the online JIT compiler, increase the performance, without sacrificing type-safety or machine-independence. Specifically, we describe mechanisms to reduce the cost of online register allocation through offline analysis, and to use offline escape analysis so the online JIT compiler can produce less heap allocations. 1 The SafeTSA Representation SafeTSA[ADvRF01] is a verifiable type safe intermediate representation designed as a machine independent intermediate representation which is both trivial to verify and easy to translate into optimized machine code. SafeTSA achieves this through the novel combination of several key features: the control structure tree, instructions in SSA form, dominator-based referential integrity, type safety through type separation, a type system extended to support key optimizations, and a carefully chosen instruction set. Figure 2 contains a graphical depiction of the safeTSA representation produced from the source program in Fig. 1. It will be referred to in the following discussion of safeTSA’s key features. Rather than allowing arbitrary branch instructions, safeTSA conveys the program’s control flow through a tree of high-level control structure elements closely paralleling those of the Java source language. This control structure tree can be seen depicted as the large font boxes and the connecting lines in Fig. 2; as a first approximation, the control structure tree can be thought of as a method’s abstract syntax tree with its expressions removed. The use of control structure trees restricts safeTSA methods’ control flow graphs to well defined