Abstract:
This research proposes a hybrid approach for implementing performance-oriented compiler intrinsics. Compiler intrinsics are special functions that provide low-level functionality and performance improvements in high-level languages. Current implementations typically use either in-place expansion or call-based methods. In-place expansion can create excessive code size and increase compile time but it can produce more efficient code in terms of execution time. Call-based approaches can lose at performance due to call instruction overhead but win at compilation time and code size. We survey intrinsics implementation in several modern virtual machine compilers: HotSpot Java Virtual Machine, and Android RunTime. We implement our hybrid approach in the LLVM-based compiler of Ark VM. Ark VM is an experimental bytecode virtual machine with garbage collection and dynamic and static compilation. We evaluate our approach against in-place expansion and call approaches using a large set of benchmarks. Results show the hybrid approach provides considerable performance improvements. For string-related benchmarks, the hybrid approach is 6.8% faster compared to the no-inlining baseline. Pure in-place expansion achieves only 0.7% execution time improvement of the hybrid implementation. We explore two versions of our hybrid approach. The "untouched" version lets LLVM control inlining decisions. The "heuristic" approach was developed after we observed LLVM's tendency to inline code too aggressively. This research helps compiler developers balance execution speed with reasonable code size and compile time when implementing intrinsics.
Fulltext:
PDF file (980 kB)