Inter-iteration Scalar Replacement Using Array SSA Form

In this paper, we introduce novel simple and efficient analysis algorithms for scalar replacement and dead store elimination that are built on Array SSA form, a uniform representation for capturing control and data flow properties at the level of array or pointer accesses. We present extensions to the original Array SSA form representation to capture loop-carried data flow information for arrays and pointers. A core contribution of our algorithm is a subscript analysis that propagates array indices across loop iterations. Compared to past work, this algorithm can handle control flow within and across loop iterations and degrade gracefully in the presence of unanalyzable subscripts. We also introduce code transformations that can use the output of our analysis algorithms to perform the necessary scalar replacement transformations (including the insertion of loop prologues and epilogues for loop-carried reuse). Our experimental results show performance improvements of up to 2.29× relative to code generated by LLVM at -O3 level. These results promise to make our algorithms a desirable starting point for scalar replacement implementations in modern SSA-based compiler infrastructures such as LLVM.