Compilation Order Matters

At design time, the compiler writer selects a set of optimizations and an order in which to apply them. These choices have a direct impact on the quality of code that the compiler can generate. They also determine, to a great extent, the set of programs for which the compiler generates good code. Making good choices is difficult for several reasons. The sheer number of transformations that have been described in the literature is daunting. The relationship between features of the input program and improvements (or degradations) produced by a transformation is not formally characterized. The interactions between these transformations are poorly understood. Finally, many of the existing transformations have overlapping effects, either singly or in combination. We have built a prototype adaptive compiler that changes its selection of transformations and its application order based on the input program, the performance characteristics of the target machine, and an explicit external objective function. This adaptive compiler performs a series of experiments to find a sequence of optimizations that minimizes the objective function. This prototype compiler lets us explore the impact of transformation selection and ordering on the code that the compiler produces. This paper describes our prototype adaptive compiler. It presents experimental results that demonstrate the impact of program specific optimization sequences. It begins to characterize the search space in which the adaptive compiler operates.