The optimization of short sequences of loop-free, fixed-point assembly code sequences is an important problem in high-performance computing. However, the competing constraints of transformation correctness and performance improvement often force even special purpose compilers to produce sub-optimal code. We show that by encoding these constraints as terms in a cost function, and using a Markov Chain Monte Carlo sampler to rapidly explore the space of all possible code sequences, we are able to generate aggressively optimized versions of a given target code sequence. Beginning from binaries compiled by 11vm –O0, we are able to produce provably correct code sequences that either match or outperform the code produced by qcc –O3, icc –O3, and in some cases expert handwritten assembly.
For many application domains there is considerable value in producing the most performant code possible. However, the traditional structure of a compiler's optimization phase is ill-suited to this task. Factoring the optimization problem into a collection of small subproblems that can be solved independently—although suitable for generating consistently good code—can lead to sub-optimal results. In many cases, the best possible code can only be obtained through the simultaneous consideration of mutually dependent issues such as instruction selection, register allocation, and target-dependent optimization.
No entries found