https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123501
--- Comment #1 from Robin Dapp <rdapp at gcc dot gnu.org> --- Another LMUL corner-case :) When the mask vector has more than 64 elements (LMUL4 is enough to trigger a miscompile) we wrongly set its nonzero bit mask to -127. The nonzero bits code seems to assume that a popcount:DI always has a DImode operand but we have e.g. (popcount:DI (V128BImode)). Testing a patch.
