Difference between revisions of "Work Package 2"
(Created page with "== Description == :Design a peephole optimization to integrate a superoptimizer into GCC and/or LLVM. == Notes == A peephole optimizer is no longer relevant - not enough po...")
Revision as of 17:09, 1 August 2014
- Design a peephole optimization to integrate a superoptimizer into GCC and/or LLVM.
A peephole optimizer is no longer relevant - not enough potential for gains.
Optimizing simultaneous sequences.
- Peephole superoptimizer study could bruteforce multiple sequences simultenously, by hashing the test vectors
- Unclear to see how this can be done for SMT based methods
- Naively attempting to find multiple solutions at the same time will limit the 'branch pruning' of the solver – it will be unable to rule out large areas of the search space due to multiple divergant solutions
- Could be the thing that limits the effectiveness of the SMT method
- Possible solutions
- Attempt to find multiple solutions to 'similar' sequences. Based on the assumption that the solutions will be similar. How to define a similarity metric?