Real-World Loops Are Easy to Predict

Abstract

The Trip Count of a loop determines how many iterations this loop performs. Several compiler optimizations yield greater benefits for large trip counts, and are either innocuous or detrimental for small ones. However, predicting exactly the trip count of a loop is an undecidable problem in general. Thus, such problem is usually approached through heuristics, which tend to be computationally expensive. In this paper we argue that in most cases there is no need to resort to expensive methods and that in many cases the trip count prediction does not need to be sound. In that sense, we propose a lightweight trip count prediction heuristic. Our method identifies the pattern on which the induction variables of each loop are updated between two iterations and generate symbolic expressions that represent the trip counts of the loops. Such expressions can be evaluated at runtime with O(1) complexity and allow blocks of code to be conditionally executed, depending on the expected trip count. We argue that such technique is useful for speculative optimizations, very common in the world of just-in-time compilers. For instance, if we predict that a loop will iterate for a long time, we can perform more aggressive JIT optimizations. Furthermore, we show that despite the simplicity of our technique, we have accurately predicted nearly 90% of all the interval loops found in millions of lines of C code. The interval loops represent approximately 67% of the total number of loops of the programs.