https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123414

--- Comment #4 from Robin Dapp <rdapp at gcc dot gnu.org> ---
Somewhere in the vicinity of vcompress the problem surely is.  The vcompresses
are part of a reduction by vector right shifts.  For those to work we must
preserve the neutral element 1 even in the elements that are not part of the
multiplication before (we have a VL of 5 and a VLEN of 256...).  We don't,
however, but have zeros instead.  As the reduction operates on full vectors
those zeros just zero out our multiplication, returning the value 0 eventually.

Need to dig a bit deeper still, though.

Reply via email to