https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127909
--- Comment #5 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #4)
> (In reply to Tamar Christina from comment #3)
> > (In reply to Richard Biener from comment #1)
> > > There is a duplicate PR for this, as dot_prod did not define the actual
> > > lane
> > > reducing order for the optabs SLP could not recognize this. ISTR we have
> > > fixed this to some extent, so this could now be leveraged as a SLP
> > > pattern.
> > >
> > > ISTR the duplicate is about SAD V16QI -> V4SI with SSE.
> >
> > Isn't this case slightly different though?
> > It probably would have worked had V1XX vectors been valid vectors for BB
> > SLP.
> >
> > But BB SLP rejects any V1 vector type.
> >
> > The above would reduce V2HI -> V1SI in vect_build_slp_tree_1
> >
> > with
> >
> > if (is_a <bb_vec_info> (vinfo)
> > && known_le (TYPE_VECTOR_SUBPARTS (vectype), 1U))
> > {
> > if (dump_enabled_p ())
> > dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> > "Build SLP failed: not using single lane "
> > "vector type %T\n", vectype);
> > matches[0] = -1;
> > return false;
> > }
> >
> > which I think is the reason this is failing now, but not sure why the
> > limitation is there for BB SLP.
>
> Because we don't want to vectorize scalar code with just vector<1>s.
> I'm not sure why you'd need those?
Well.. specifically for lane reducing operations from N lanes to 1?