Universal Instruction Selection


Gabriel Hjort Blindell (now Gabriel Hjort Åkerlund)


KTH Royal Institute of Technology


Christian Schulte
Mats Carlsson
Ingo Sander


In code generation, instruction selection chooses instructions to implement a given program under compilation, global code motion moves computations from one part of the program to another, and block ordering places program blocks in a consecutive sequence. Local instruction selection chooses instructions one program block at a time while global instruction selection does so for the entire function. This dissertation introduces a new approach called universal instruction selection that integrates global instruction selection with global code motion and block ordering. By doing so, it addresses limitations of existing instruction selection techniques that fail to exploit many of the instructions provided by modern processors.

To handle the combinatorial nature of these problems, the approach is based on constraint programming, a combinatorial optimization method. It relies on a novel model that is simpler and more flexible compared to the techniques used in modern compilers and that captures crucial features ignored by other combinatorial approaches. The dissertation also proposes extensions to the model for integrating instruction scheduling and register allocation, two other important problems of code generation.

The model is enabled by a novel, graph-based representation that unifies data and control flow for entire functions. The representation is crucial for integrating instruction selection with global code motion and for modeling sophisticated instructions, whose behavior contains both data and control flow, as graphs.

Through experimental evaluation, universal instruction selection is demonstrated to handle architectures with a rich instruction set and scales up to functions with hundreds of operations. For these functions, it generates code of equal or better quality compared to the state of the art. The dissertation also demonstrates that there is sufficient data parallelism to be exploited through selection of SIMD instructions and that this exploitation benefits from global code motion. With these results, it is argued that constraint programming is a flexible, practical, competitive, and extensible approach for combining global instruction selection, global code motion, and block ordering.


Friday, April 6, 2018

PDF of thesis: