Pointer analysis is widely used as a base for different kinds of static analyses and compiler optimizations. Designing a scalable pointer analysis with acceptable precision for use in production compilers is still an open question. Modern object oriented languages like Java and Scala promote abstractions and code reuse, both of which make it difficult to achieve precision. Collection data structures are an example of a pervasively used component in such languages. But analyzing collection implementations with full context sensitivity leads to prohibitively long analysis times. We use semantic models to reduce the complex internal implementation of, e.g., a collection to a small and concise model. Analyzing the model with context sensitivity leads to precise results with only a modest increase in analysis time. The models must be written manually, which is feasible because a model method usually consists of only a few statements. Our implementation in GraalVM Native Image shows a rise in useful precision (1.35X rise in the number of checkcast statements that can be elided over the default analysis configuration) with a manageable performance cost (19% rise in analysis time). CCS Concepts • Software and its engineering → Automated static analysis.

Batching has a fundamental influence on the efficiency of deep neural network (DNN) execution. However, for dynamic DNNs, efficient batching is particularly challenging as the dataflow graph varies per input instance. As a result, state-ofthe-art frameworks use heuristics that result in suboptimal batching decisions. Further, batching puts strict restrictions on memory adjacency and can lead to high data movement costs. In this paper, we provide an approach for batching dynamic DNNs based on finite state machines, which enables the automatic discovery of batching policies specialized for each DNN via reinforcement learning. Moreover, we find that memory planning that is aware of the batching policy can save significant data movement overheads, which is automated by a PQ tree-based algorithm we introduce. Experimental results show that our framework speeds up state-of-the-art frameworks by on average 1.15x, 1.39x, and 2.45x for chain-based, tree-based, and lattice-based DNNs across CPU and GPU.

No abstract