https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120805
--- Comment #19 from Tamar Christina <tnfchris at gcc dot gnu.org> --- (In reply to Avinash Jayakar from comment #18) > (In reply to Tamar Christina from comment #17) > > (In reply to Avinash Jayakar from comment #16) > > > Created attachment 61956 [details] > > > I have just changed the order within the conditional where the const_vf > > > gets > > > assigned, to have it behave in similar way as before (use of log_vf) > > > > That to me is just a plaster rather than a fix. You've flipped the > > initialization of logvf rather than just adding the explicit requirement > > that this code can't support epilogues. > > ok, but adding a statement that it does not support epilogue is also a > plaster right? because there can be other cases where > LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo) > is true, and epilogue is just one of them. So if there is a scenario like > that, then we would have to not support those cases explicitly. > If we are definitely sure that LOOP_VINFO_USING_PARTIAL_VECTORS_P > (loop_vinfo) would be true for epilogues only when vf > 0, then the fix you > suggested will work. I am not sure if that is the case. If the patch is not > convincing, you can take this issue. But partial loops are *not* the problem. LOOP_VINFO_USING_PARTIAL_VECTORS_P just means your loop is masked, i.e. you may have a partial iteration. If LOOP_VINFO_USING_PARTIAL_VECTORS_P is on the main loop the range *is* correct because niters_vector is then an actual vector iteration count. You can have a partial main loop and zero epilogues. This is usually just a cost modeling decision. Epilogues are the problem, and before the check on !LOOP_VINFO_USING_PARTIAL_VECTORS_P just happen to block epilogues on PowerPC because PowerPC is a non-poly target which generated an unpredicated main loop and a predicated non-VLA epilogue. It's when it's an epilogue that niters_vector has not been adjusted to vector iterations. i.e. it's scalar. that's why you cannot set a vector iteration bounds on the scalar bnd value as I explained here https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120805#c13. When it's an epilogue, the bounds value is not used in the latch of the loop, because there is no latch, epilogues will have their loop control removed as they can by definition only do VF values (i.e, one iteration).
