On Wed, May 15, 2013 at 7:02 AM, Jeff Law <l...@redhat.com> wrote: > On 05/14/2013 03:14 PM, Steve Ellcey wrote: >> >> >> While Jeff works on the threader, I was wondering if I could get approval >> for >> just the dominance.c part of the patch. This would allow me to use my >> pass as >> a dynamically loaded optimization pass. But without this change to >> dominance.c, >> the compiler aborts in iterate_fix_dominators when I do that. >> >> Steve Ellcey >> sell...@imgtec.com >> >> >> >> 2013-05-14 Steve Ellcey <sell...@imgtec.com> >> >> * dominance.c (iterate_fix_dominators): Add null check. > > I'd like to understand this a little more before we go forward with it. > > AFAICT, that routine is trying to incrementally update the dominators using > knowledge that the region you've copied is SESE. It's unclear what happens > in the region is not SESE. > > Threading mucks up the dominator tree in fairly serious ways and to the best > of my knowledge neither of the calls to thread across edges make any attempt > to incrementally update the dominator tree. They wipe it completely, they > also have to be quite careful in how they manipulate the various graphs to > avoid getting into an inconsistent state, then calling routines that assume > consistent state. > > > I realize you're trying to do the same, but by using the SESE copier, you're > implicitly trying to do an incremental update. I think you're going to > really need to look at the assumptions of that code and verify that the > switch FSA optimization doesn't violate those assumptions.
Indeed. I'd rather have a flag to the SESE copier that tells it the region is SEME and in that case make it not update dominators but leave that to the caller (which means, recompute them). It seems the code already handles ME regions if the other exits are "not important" (we don't have to insert/update PHI nodes in those exit blocks), so it may be that there are even more constraints on those unimportant exits due to the iterative dominator update - I think that the entry block of the region needs to dominate all exit destinations, otherwise get_dominated_by_region is not working correctly. In your case one exit is a back-edge? We should be able to add checking to the code to verify unimportant edges are unimportant "enough". I'm sure Zdenek knows the limitations best. Richard. > Jeff