On 04/10/2017 11:37 PM, Ulrich Weigand wrote: > Dominik Vogt wrote: > >> So, we could add a special case for const0_rtx that generates the >> LT pattern and does not rely on Combine, and get rid of the >> peephole. I'm not sure this is worthwhile thoug, because the >> peephole has other beneficial effects (as discussed), and until >> we've solved the problems preventing Combine from merging L+LTR in >> some cases, this is the best we have. What do you think? > > If we removed the peephole (for now), the patch now only touches > parts of the backend used to emit atomic instructions, so code > generation for any code that doesn't use those is guaranteed to > be unchanged. Given that we're quite late in the cycle, this > might be a good idea at this point ... > > But I don't see anything actually incorrect in the peephole, and > it might indeed be a good thing in general -- just maybe more > appropriate for the next stage1. > > Andreas, do you have an opinion on this?
Yes, the change is unrelated to the rest of the patch and therefore should go into a separate patch. But I'm fine with applying both right now. The peephole surprisingly often helped code generation. There appear to be plenty of cases where the L+LTR combination could not be handled otherwise. Dominik reviewed the speccpu diffs and there were only improvements. -Andreas-
