On Sat, Nov 02, 2019 at 10:14:15AM -0600, Jeff Law wrote: > On 11/2/19 9:51 AM, Georg-Johann Lay wrote: > > Segher Boessenkool schrieb: > >>> Btw, does GCC support clobbering registers in branches (or > >>> cbranch<mode>4 for that matter)? This requirement would come up when > >>> transitioning avr to cc_mode because cbranches would live post reload. > >> > >> Of course. You cannot have *reloads* on branches, that is all.
Sigh. *Output reloads*, sorry for dropping a crucial word here :-/ > > Does this also apply to input reloads? > > This should work in the LRA world just fine. The problems are with > output reloads, not input reloads. In the old reload world as well. Only output reloads are a problem. > With an output reload on a branch you have to emit the reload at the > target(s) of the branch as well as the fallthru path (if one exists). Or you can do something *before* the branch, but that is different for every case, and harder to do anyway. If you insert the reloads at the jump target, you'll need to make a separate predecessor of the jump target and insert the reload *there*, if we aren't the only predecessor of that target. > In theory we could do this for simple branches, but for indirects it'd > be virtually impossible. Yup. Luckily most branches with outputs are not indirect branches. I would love if this problem was fixed once and for all (I only care about LRA here of course). Segher
