Scalar replacement is an effective optimization for removing memory accesses. However, exposing all possible array reuse with scalars may cause a significant increase in register pressure, resulting in register spilling and performance degradation. In this paper, we present a low cost method to predict the register pressure of a loop before applying scalar replacement on high-level source code, called Pseudo-schedule Register Prediction (PRP), that takes into account the effects of both software pipelining and register allocation. PRP attempts to eliminate the possibility of degradation from scalar replacement due to register spilling while providing opportunities for a good speedup.

PRP uses three approximation algorithms: one for constructing a data dependence graph, one for computing the recurrence constraints of a software pipelined loop, and one for building a pseudo-schedule. Our experiments show that PRP predicts the floating-point register pressure within 2 registers and the integer register pressure within 2.7 registers on average with a time complexity of O(n^2) in practice. PRP achieves similar performance to the best previous approach, having O(n^3) complexity, with less than one-fourth of the compilation time on our test suite.