- This project implements several improvements for automatic prefetching of array accesses in loops:
- Cross-loop reuse analysis
- Peeling of loops to enable better handling of reuses
- Nontemporal store generation
- Tuning of prefetching parameters and heuristics, and adding more parameters to control them
- Zdenek Dvorak
- Most of the code is written and tested for correctness; however, no serious performance tuning has been done yet. This should be finished by October 30th, 2006.
- Performance gains with -fprefetch-loop-arrays, especially for stream processing code.
- The changes can be roughly split to the parts as noted in the project description. Cross-loop reuse analysis is a completely new analysis pass, the remaining parts are basically minor changes to the existing prefetching/loop manipulation frameworks.