https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121910
--- Comment #6 from JuzheZhong <juzhe.zhong at rivai dot ai> --- (In reply to Zhongyao Chen from comment #5) > I have doubts about the following code: > Why does it return regno_alignment when the calculated lmul is greater than > RVV_M8? > > I change the return to RVV_M8, and then the issue disappeared, no register > spill. However, I'm unsure if the changes is appropriate. what you think? > > > @@ -632,7 +726,7 @@ compute_estimated_lmul (loop_vec_info loop_vinfo, > machine_mode m > ode) > int estimated_lmul = estimated_vf * GET_MODE_BITSIZE > (mode).to_constant () > / TARGET_MIN_VLEN; > if (estimated_lmul > RVV_M8) > - return regno_alignment; > + return RVV_M8; > else > return estimated_lmul; > } The changes look good to me. Could you try the GCC testsuite testing with your change ? It looks reasonable to return maximum LMUL (RVV_M8) instead of regno_alignment.
