On 16 Nov 2014, at 22:18, Segher Boessenkool <seg...@kernel.crashing.org> wrote:
> On Sun, Nov 16, 2014 at 05:45:06PM +0900, Oleg Endo wrote: >> When you commit those, could you please also add PR 59278 to the ChangeLog so >> that the commit appears in bugzilla? After your patches are in, I'd like to >> add some SH specific test cases (assuming that your patches fix PR 59278). > > It doesn't fix this testcase. Too bad. > Here, recog_for_combine needs to add a > clobber of T, but it thinks T is not dead. It doesn't say anything > about that in the debug dump :-( > > Maybe the movrt_negc pattern shouldn't set T at all, just clobber it? On SH, it's not just that particular pattern, but a couple of others, which would need to be changed from set-set to set-clobber before/during combine and then converted/split into the actual set-set patterns after combine. E.g. some patterns set the T bit to a known zero/one value which can be good to know later on. > Or reg_dead_at_p can be taught about "unused" notes. Sounds the easier way from my point of view. I don't know about side effects for other targets of "unused" reg notes are treated as "dead" reg notes. Cheers, Oleg