Semantics for Transactional Languages Michael Scott, U Rochester Within a few years many -- perhaps most -- multicore processors are likely to be equipped with some form of hardware transactional memory, and TM extensions are likely to be wanted in multiple programming languages. Those of us working in the area have a brief opportunity to "get the semantics right." In my talk I will attempt to summarize my thoughts on the codification of TM semantics, with a particular emphasis on - the centrality of atomicity (as opposed to locking) - the non-issue of strong atomicity - the need to make confine speculation to the level of implementation - the need to consider not only opacity but also sandboxing as implementation options - the opportunity to integrate atomicity with emerging work on "deterministic parallel programming"