https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107986
--- Comment #7 from Andrew Macleod <amacleod at redhat dot com> --- (In reply to Richard Biener from comment #3) > Note the old VRP was able to at least tell that qa was [0, +INF] by means of > iterating and saturating to +INF after too many iterations. SCEV isn't of > help here. Since we've lost the last lattice-based (iterating) VRP pass now > what's the plan for cases like this? When we first process a statement we set the LHS to varying first. Then we resolve all the inputs, and then do the final calculation of the LHS and save that. This is done to prevent resolving inputs along back edges from entering a cycle. If the back edge ever returns to this def, it will pick up the saved varying value and stop looking.. This is conservatively correct, but it does mean in the case of PHI nodes, we are not optimistic about the outcome if there is a cyclic dependency. I intend to experiment with pre-walking any back edges to look for bounds if there is a cycle.. ie, the range increases or decreases providing an upper/lower limit. I just haven't gotten to it yet. SO no iterations as such, but a quick pattern check which may provide a bound. in another PR (107972) I mentioned al alternative VRP which discovers things from the bottom to the top.. Theres some possibility that pass might expose this sort of thing too when I get to it.
