Parallelization Based on Strict Independence

Conclusions

• Algorithms
  • Greedy strategies (CDG) are able to exploit more parallelism
    although at a cost in compile time in some cases
  • Exploiting only unconditional parallelism (UDG) is a good fall-back strategy
    but only if accurate analyses are used
  • Simple heuristics (MEL) are effective as a fall-back strategy for CDG in many cases
    but (good) opportunities are missed in several others
  • In general: use a ``bounded'' CDG which falls back to MEL-like heuristics or unconditional strategies
    proposed UCDG

• Analyses
  • Groundness no dependencies exist
  • Sharing dependencies can be tracked
  • Linearity definite independence
    Via accurate propagation of non-sharing
  • Freeness definite dependence and independence
    Via non-groundness, definite aliasing, and (restricted) linearity
  • Combination of all found best in results and quite inexpensive in time
  • Quite effective overall ...and feasible!