Improvements to graph coloring register allocation

Reviewer: Rajiv Gupta

The authors identify problems with the Chaitin-style graph coloring register allocator and propose solutions to these problems. The first problem is unnecessary spilling of values. Chaitin's allocator continues to spill values until the resulting interference graph is guaranteed to be k -colorable, where k is the number of registers. Registers may in fact be available for some of the spilled values, however. The authors develop an optimistic coloring algorithm that attempts to assign registers to values that would otherwise be spilled by Chaitin's algorithm. The second problem addressed is reducing the cost of spill code in situations where the spill code must be generated. In some situations, it is cheaper to recompute the value than to store it and retrieve it from memory. In particular, if a constant value must be spilled, it is cheaper to regenerate the value using a single immediate load instruction than to store and load the value. The authors present a technique called rematerialization to take advantage of such situations. The experimental results suggest that execution time can be reduced by up to 40 percent using optimistic coloring, and up to 20 percent improvement can result using rematerialization. A thorough comparison with Chow's priority-based register allocator [1] would have been valuable. Since Chow's algorithm does not attempt to guarantee the colorability of the interference graph through repeated spilling, it is unlikely to suffer from the problem of unnecessary spilling of values. Chow's algorithm can also benefit from the opportunities exploited through rematerialization, however. The paper is well written and provides an excellent summary of Chaitin's register allocator as well as valuable insight into the implementation of such register allocators. In addition, some problems with Chaitin's allocator are identified, and modifications that successfully address these problems are developed.