Abstract interpretation is widely accepted as a natural framework for semantics-based analysis of program properties. However, most formulations of abstract interpretation are in terms of high-level semantic entities that do not adequately address the needs of low-level optimizations. In this paper we discuss the role of abstract interpretation in low-level compiler optimizations, examine some of its limitations, and consider ways in which they might be addressed.