Automatonlike Generation of Peephole Superoptimizers
Robotic Generation of Peephole Superoptimizers , Sorav Bansal and Alex Aiken, ASPLOS 2006.
Peephole optimizers are typically made expending human-saved pattern equaling rules, an approach that involves expertise and time, as good as being less than taxonomical at taping all opportunities for optimization. We research fully machinelike construction of peephole optimizers expending brute force superoptimization. While the optimizations discovered by our reflex system may be less universal than human-spelt counterparts, our approach has the potential to mechanically memorise a database of thousands to millions of optimizations, in contrast to the hundreds found in current peephole optimizers. We picture through an experiment that our optimizer is capable to tap performance opportunities not found by subsisting compilers; in especial, we depict speedups from 1.7 to a factor of 10 on some figure intensive kernels over a established optimising compiler.
It’s e’er fun to picture a method that ought to be intractable shown amenable through a combination of cleverness and strategically put on brute force.
I saw their performance measurements implicative but puzzling. Their results on their kernels are really astoundingly well, which they arrogate that is because they work use of the x86’s SIMD instructions. But there is very slight change in the SPEC benchmarks, and they later indicate that their peephole optimizer gets many things that other compilers get via dataflow optimization.
