On Mon, Mar 25, 2019 at 11:39:27AM -0600, Jeff Law wrote: > On 3/24/19 11:42 AM, Segher Boessenkool wrote: > > On Fri, Mar 22, 2019 at 12:16:30PM -0600, Jeff Law wrote: > >> So I finally started looking at the fpr-moves regression in this BZ. No > >> surprise this is further fallout from the combiner changes. > > > > It *exposed* the problem, yes :-) > Well, it was one of regressions that were a result of that patch. I'm a > bit disappointed in how many were just punted rather than really > analyzed and either mitigated or explicitly moved out to gcc-10 because > mitigation was too painful with marginal benefits to do at this stage. > > >> One might reasonably ask if combine's avoidance of > >> hard regs should be loosened for consecutive insns -- combining > >> consecutive insns isn't going to increase the live range of these hard > >> regs. > > > > But that is not the only reason we don't want to forward hard registers. > > One of the important reasons is correctness: we could create situations > > that reload/LRA cannot handle (cannot *possibly* handle). Another reason > > is that combine should not try to do register allocation's job: we on > > average get better code after the combine hard reg change. > No doubt. My point was that we can and should have been looking for > ways to mitigate the fallout. Which in my mind includes further > investigation of whether or not we've hit the right balance for what > should and should not be combined.
Do you have some example of something that happened after the change, that did not happen before the change (when using a var instead of a function argument, say)? > > For splitters after reload you have to do a lot of work manually, to get > > reasonable code. This is a problem everywhere :-( > Yup. In fact, we've seen multiple BZs this cycle where a DCE pass right > before/after splitting would help. In both cases we ended up hacking up > the splitters a bit. It's one of the items I think we'll want to > revisit during gcc-10 stage1 development. I would like to even see something CSE-like, cprop-like, and combine-like. But most of all, people should stop doing splitters after reload if there is any way to avoid it! Segher
