On Mon, 21 Jan 2019, Richard Sandiford wrote:
> Richard Biener <rguent...@suse.de> writes:
> > @@ -5742,8 +5740,14 @@ vectorize_fold_left_reduction (stmt_vec_info
> > stmt_info,
> > auto_vec<tree> vec_oprnds0;
> > if (slp_node)
> > {
> > - vect_get_vec_defs (op0, NULL_TREE, stmt_info, &vec_oprnds0, NULL,
> > - slp_node);
> > + auto_vec<vec<tree> > vec_defs (2);
> > + auto_vec<tree> sops(2);
> > + sops.quick_push (ops[0]);
> > + sops.quick_push (ops[1]);
> > + vect_get_slp_defs (sops, slp_node, &vec_defs);
> > + vec_oprnds0.safe_splice (vec_defs[1 - reduc_index]);
> > + vec_defs[0].release ();
> > + vec_defs[1].release ();
> > group_size = SLP_TREE_SCALAR_STMTS (slp_node).length ();
> > scalar_dest_def_info = SLP_TREE_SCALAR_STMTS (slp_node)[group_size -
> > 1];
> > }
>
> Ewww. :-) Would be nice if it was easier to do the right thing.
Working on that ;)
> But are you sure we want this for fold-left reductions? The other
> operand isn't supposed to be vectorised -- it stays a scalar even
> in the vector loop.
It seems to "work" for me. The issue appears with MINUS_EXPR
fold-left reductions (you may have guessed that). I guess
it's possible to create a SLP testcase even without patched trunk
(I have group_size == 1 SLP instances "enabled").
In the end vect_get_slp_defs will, if no actual vectorized stmts
are recorded, not push any to the defs vector which means
vec_defs[0] will remain empty.
Richard.
