Partial evaluators can be separated into two classes: offline specializers, which make all of their reduce/residualize decisions before specialization, and online specializers, which make such decisions during specialization. The choice of which method to use is driven by a tradeoff between the efficiency of the specializer and the quality of the residual programs that it produces. Existing research describes some of the inefficiencies of online specializers, and how these are avoided using offline methods, but fails to address the price paid in specialization quality. This paper motivates research in online specialization by describing two fundamental limitations of the offline approach, and explains why the online approach does not encounter the same difficulties.

The degree to which a partial evaluator can specialize a source program depends on how accurately the partial evaluator can represent and maintain information about runtime values. Partial evaluators always lose some accuracy due to their use of finite type systems; however, existing partial evaluation techniques lose information about runtime values even when their type systems are capable of representing such information. This paper describes two sources of such loss in existing specializers, solutions for both cases, and the implementation of these solutions in our partial evaluation system, FUSE.