https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109971
--- Comment #6 from JuzheZhong <juzhe.zhong at rivai dot ai> --- >> With decrement IV, the optimized IR actually becomes better, it also aligns >> >>with our discussion here: >> https://gcc.gnu.org/pipermail/gcc-patches/2023->>April/615629.html (Thanks >> for the improvement!) Oh, I see. It seems that my patch is overall beneficial to PowerPC? That's what I want since I want my patch can not only help RVV but also other targets. >>It exposes something inefficient at -O2, it seems we can teach cunroll >>further >>about this kind of new sequence. Ok, we can optimize it for decrement IV in the future. >>If you meant to disable decrement IV on Power (but now actually enable it >>>>unexpectedly), then probably we can just keep it (not disabling), for Power >>we >>mainly adopt --param=vect-partial-vector-usage=1, it shouldn't be >>affected, for >>--param=vect-partial-vector-usage=2, it does generate better >>code sequence for >>most cases and we can improve the remaining worse one >>gradually. I am not meant to disable decrement IV on Power. Actually, I really hope Power can reuse the flow that I build for RVV. It makes things more meaningful. If it works for power and it seems to improve power codegen in most cases, I'd like to see power enable it by default. Then we can optimize it togther. Thanks a lot for your information.
