Tailoring Graph-coloring Register Allocation For Runtime Compilation

Abstract

Just-in-time compilers are invoked during application execution and therefore need to ensure fast compilation times. Consequently, runtime compiler designers are averse to implementing compile-time intensive optimization algorithms. Instead, they tend to select faster but less effective transformations. In this paper, we explore this trade-off for an important optimization -global register allocation. We present a graph-coloring register allocator that has been redesigned for runtime compilation. Compared to ChaitinBriggs [7], a standard graph-coloring technique, the reformulated algorithm requires considerably less allocation time and produces allocations that are only marginally worse than those of Chaitin-Briggs. Our experimental results indicate that the allocator performs better than the linear-scan and Chaitin-Briggs allocators on most benchmarks in a runtime compilation environment. By increasing allocation efficiency and preserving optimization quality, the presented algorithm increases the suitability and profitability of a graph-coloring register allocation strategy for a runtime compiler.

Performance Comparison of Register Allocation Algorithms in Dynamic Binary Translation

“…The allocator outperforms both Chaitin-Briggs and linear-scan allocators for most benchmarks in a runtime compilation environment. On average, it improved application performance by 9% over Chaitin-Briggs and 7% over linear-scan [22].…”

“…In practice, the graphcoloring approach is more proficient at allocating registers than linear-scan techniques and can improve the performance of the allocated code [20] [21]. In a JIT environment, the additional compile time required by the algorithm greatly diminishes the runtime gains achieved by an improved allocation [22]. The interference graph builder is the most expensive component-its worst-case asymptotic bound is O(n 2 ), where n is the number of live ranges in the program [22].…”

“…In a JIT environment, the additional compile time required by the algorithm greatly diminishes the runtime gains achieved by an improved allocation [22]. The interference graph builder is the most expensive component-its worst-case asymptotic bound is O(n 2 ), where n is the number of live ranges in the program [22]. The interference graph builder is the most expensive component.…”

“…The interference graph builder is the most expensive component. On average, it consumes about 72% of the allocation process [22].…”

“…lossy [22] is a graph-coloring register allocator that has been redesigned for runtime compilation. It constructs the interference graph once and then uses incremental methods to update the graph after spilling and coalescing.…”