Convergent Compilation Applied to Loop Unrolling

  • Authors:

  • Nicholas Nethercote;Doug Burger;Kathryn S. Mckinley

  • Affiliations:

  • The University of Texas at Austin,;The University of Texas at Austin,;The University of Texas at Austin,

  • Venue:
  • Transactions on High-Performance Embedded Architectures and Compilers I
  • Year:
  • 2007

Quantified Score

Hi-index 0.00

Visualization

Abstract

Well-engineered compilers use a carefully selected set of optimizations, heuristic optimization policies, and a phase ordering. Designing a single optimization heuristic that works well with other optimization phases is a challenging task. Although compiler designers evaluate heuristics and phase orderings before deployment, compilers typically do not statically evaluate nor refine the quality of their optimization decisions during a specific compilation.This paper identifies a class of optimizations for which the compiler can statically evaluate the effectiveness of its heuristics and phase interactions. When necessary, it then modifies and reapplies its optimization policies. We call this approach convergent compilation, since it iterates to converge on high quality code. This model incurs additional compilation time to avoid some of the difficulties of predicting phase interactions and perfecting heuristicsThis work was motivated by the TRIPS architecture which has resource constraints that have conflicting phase order requirements. For example, each atomic execution unit (a TRIPS block) has a maximum number of instructions (128) and a fixed minimum execution time cost. Loop unrolling and other optimizations thus seek to maximize the number of mostly full blocks. Because unrolling enables many downstream optimizations, it needs to occur well before code generation, but this position makes it impossible to accurately predict the final number of instructions. After the compiler generates code, it knows the exact instruction count and consequently if it unrolled too much or too little or just right. If necessary, convergent unrolling then goes back and adjusts the unroll amount accordingly and reapplies subsequent optimization phases. We implement convergent unrolling which automatically matches the best hand unrolled version for a set of microbenchmarks on the TRIPS architectural simulator.Convergent compilation can help solve other phase ordering and heuristic tuning compilation challenges. It is particularly well suited for resource constraints that the compiler can statically evaluate such as register usage, instruction level parallelism, and code size. More importantly, these resource constraints are key performance indicators in embedded, VLIW, and partitioned hardware and indicate that convergent compilation should be broadly applicable.