Compilers often use escape analysis to elide locking operations on thread-local data. Similarly, dynamic race detectors may use escape analysis to elide race checks on thread-local data. In this paper, we study the correctness of these two related optimizations when using a partial escape analysis, which identifies objects that are currently thread-local but that may later become thread-shared.

We show that lock elision based on partial escape analysis is unsound for the Java memory model. We also show that race check elision based on a partial escape analysis weakens the precision of dynamic race detectors. Finally, we prove that race check elision based on a partial escape analysis is sound with respect to this weakened, but still useful, notion of precision.