Automatic Generation of Peephole Superoptimizers

Sorav Bansal and Alex Aiken

Gokulan R

CS15B033
29 April 2020

Summary of the paper

Peephole optimization: Replace one instruction sequence with a better instruction sequence
Goal: Automatically build sequences instead of hand-tuned optimizations
Three-part optimizer:
- harvester: extract instruction sequences to optimize
- enumerator: enumerates all possible sequences upto a certain length exhaustively, checking if each candidate is a replacement
- optimization database: index of all discovered optimizations to applications

Key Algorithm

offline training phase: // input: training program
    harvester: extract instruction sequences
    canonicalizer: rename registers to reduce search space
    fingerprinter: compute fingerprint for instruction sequence 
    and store in fingerprinting table

    enumerator: generate sequence of possible assembly instructions
        fingerprinting: fingerprinting of a generated sequence
        search in fingerprinting table: check if any equivalent sequence
        is present in training program
        if yes:
            boolean check: check if contexts match
            if yes:
                store in optimization table

online optimization phase: // input: program to compile
    while no optimization exists:
    	harvest an instruction sequence
        canonicalize it
        fingerprint it
        search optimization table
        replace sequence if optimization found

Future Work

Increasing the length of instruction sequence with better compute resources
Can be extended to PTX assembly code used in CUDA supported GPUs
Current work optimizes for CISC assembly - can be extended to RISC assembly instruction - MIPS, RISC-V
Efficient compression scheme for storing in optimization database
Sequence-specific testvector generation to verify correctness