Accurate static estimators for program optimization

Abstract

Determining the relative execution frequency of program regions is essential for many important optimization techniques, including register allocation,

Preference-Guided Register Assignment

“…This execution frequency can either be gathered from profile data or estimated (see e.g. [10]). For the sake of brevity, let ∈ {use, def }.…”

“…This can be done heuristically (cf. Wagner et al [10]) or they can be obtained from profiling information. Using the execution frequencies, we calculate the trace value of each block: The trace value of a block is the maximum of the trace values of its control flow predecessors (disregarding back edges) plus its own execution frequency.…”

“…The benchmarks mostly exercise the seven generalpurpose registers of the x86. The execution frequencies were statically estimated using a Markov-chain model [10]. We compare the algorithm presented in this paper with our previous work performing colaescing by recoloring [5] after register allocation.…”

Performance issues in correlated branch prediction schemes

“…Beside the obvious implications for improvements in shared libraries, the fact that library routines can be accurately predicted by profiles from different applications implies that static branch prediction methods, such as those of Ball and Larus [1], our own method [4] and related methods [14,16], can be improved. Table 15: Overall miss rates for the 17 programs common between this study and the ESP study.…”

“…However, many researchers believe that collecting profile information is too costly or time-consuming, and that many programmers may not collect such information. Thus, there has been considerable interest in heuristic prediction, or predicting the behavior of a program from the program's structure [1,4,16,14]. These methods use heuristics or statistical information to predict a programs behavior.…”