Lost in translation: Exposing hidden compiler optimization opportunities
share
To increase productivity, today's compilers offer a two-fold abstraction: they hide hardware complexity from the software developer, and they support many architectures and programming languages. At the same time, due to fierce market competition, most processor vendors do not disclose many of their implementation details. These factors force software developers to treat both compilers and architectures as black boxes. In practice, this leads to a suboptimal compiler behavior where the maximum potential of improving an application's resource usage, such as execution time, is often not realized. This paper targets all three communities, compiler engineers, software developers and hardware architects, by exposing missed optimization opportunities. By exploiting the behavior of the standard optimization levels, such as the -O3, of the LLVM v6.0 compiler, we show how to reveal hidden cross-architecture and architecture-dependent potential optimizations on two popular processors: the Intel i5-6300U, widely used in portable PCs, and the ARM Cortex-A53-based Broadcom BCM2837 used in the Raspberry Pi 3B+. The classic nightly regression testing can then be extended to use the resource usage and compilation information collected while exploiting subsequences of the standard optimization levels. This provides a systematic means of detecting and tracking missed optimization opportunities. The enhanced nightly regression system is capable of driving the improvement and tuning of the compiler's common optimizer.
READ FULL TEXT