SSA-Form-Based Register Allocation for the Java HotSpotTM Server Compiler

Register allocation, i.e., the task of assigning processor registers to local variables and temporary values, is one of the most important compiler optimizations. A vast amount of research has led to algorithms ranging from simple and fast heuristics to optimal algorithms with exponential time complexity. Because the problem is known to be NP-complete [2], algorithms must balance the time necessary for allocation against the resulting code quality. Two common algorithms in modern compilers are graph coloring (see for example [1, 2]), which is suitable when compilation time is not a major concern, and linear scan [6, 8, 10], which is faster and therefore frequently used for just-in-time compilers where compilation time adds to run time. Static single assignment (SSA) form [3] is a type of intermediate representation that simplifies many compiler optimizations. All variables have only a single point of definition. At control flow joins, phi functions are used to merge different variables of the predecessor blocks. Because processors cannot execute phi functions, it is necessary to replace them with move instructions during code generation (SSA form deconstruction). Traditionally, SSA form deconstruction was performed before register allocation. Only recently has it been observed that register allocation on SSA form has several advantages due to additional guarantees on variable lifetime. Lifetime information is essential for register allocation because two variables that interfere, i.e., that are live at the same time, must not have the same register assigned. The interference graph of a program in SSA form is chordal (every cycle with four or more edges has an edge connecting two vertices of the cycle, leading to a triangulated structure). Many graph algorithms are simpler on chordal graphs, e.g., graph coloring can be performed in polynomial time. These properties were used to simplify register allocators based on graph coloring [4]. When the maximum register pressure is below or equal to the number of available registers, allocation is guaranteed to succeed. This allows to split the algorithms for spilling and register assignment. Traditionally, spilling and register assignment were interleaved, i.e., a variable was spilled when the graph turned out to be not colorable. This led to a time-consuming repeated execution of the graph coloring algorithm. Recent research at the University of California, Irvine, also applied the benefits of SSA Form on linear scan register allocation [9]. The lifetime intervals, which are the basic data structure of the algorithm, are easier to construct and have a simpler structure. Additionally, infrastructure already present in the linear scan algorithm can be used to perform SSA form deconstruction after register allocation, thus making a separate SSA form deconstruction algorithm unnecessary. When comparing graph coloring and linear scan register allocation on SSA form, it turns out the two algorithms are not so different than they look at the first glance. The simplifications allowed by SSA form eliminate parts that were previously handled differently. Therefore, the next research step is to analyze the cases where the two algorithms are equivalent (e.g., register allocation without spilling is very similar), and find cases where the algorithms still differ (e.g., selecting variables to be spilled). Finally, both algorithms should be merged to one algorithm that covers graph coloring and linear scan as special cases, but also allows a gradual “sliding” between them, i.e., it should be configurable to what extent graph coloring and linear scan are used.