On 12/13/2017 09:24 AM, Richard Biener wrote: >> >> Alternately we could to the dom_walker ctor that an initial state of >> EDGE_EXECUTABLE is already set. > > I'm quite sure that wouldn't help for VRP. Not sure why. But it's not worth digging deep into.
I do think the current structure could still fail to pick up some secondary cases where blocks become unreachable as a result of both not needing to visit during the lattice propagation step and the substitution step. But I'd expect this to be rare. > I think David's approach is fine just we don't need any other API to get at a > known executable outgoing edge. We can improve the existing one or just add > the trivial folding required. I think Michael's suggestion to pass in NULL for the value and allow find_edge to try and determine the value makes the most sense here. Jeff
