On 26/08/13 10:00, Jan Stolarek wrote:
1. This adds a new constraint on the dataflow algorithm, namely that it must traverse the blocks in the correct order. I don't follow. From what I've seen in the code Hoopl orders blocks using depth-first traversal. For backwards analysis it reverses that order. I don't think that proposed changes will affect this. What did I miss?
You didn't miss anything. The new constraint is that forwards analysis *must* process blocks in this order, whereas previously it didn't have to.
2. Has anyone tried implementing this change? I'm slightly concerned that having a difference between forward and backward analyses might lead to divergence in some of the shared parts of the internals of the algorithm. If it turns out that forward and backward cases have to diverge to implement this change, will that be acceptable? GHC's specialized Hoopl module (compiler/cmm/Hoopl/Dataflow) already has separate functions for handling forward and backward analysis.
But they share the core fixpoint algorithm underneath: fixpoint, fixpointAnal, and some auxiliary functions. It's not a big deal. This code is heavily tuned though, so keep an eye on performance. Cheers, Simon