On Fri, Nov 17, 2023 at 9:45 AM Robin Dapp <rdapp....@gmail.com> wrote:
>
> > But note you can explicitly specify a vector type as well, there's an
> > overload for it, so we can fix the "invariant" case with the following
> > (OK if you can test this on relevant targets)
> >
> > diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> > index 3f59139cb01..936a3de9534 100644
> > --- a/gcc/tree-vect-loop.cc
> > +++ b/gcc/tree-vect-loop.cc
> > @@ -8450,12 +8450,14 @@ vect_transform_reduction (loop_vec_info loop_vinfo,
> > value. */
> > bool cond_fn_p = code.is_internal_fn ()
> > && conditional_internal_fn_code (internal_fn (code)) != ERROR_MARK;
> > + int cond_index = -1;
> > if (cond_fn_p)
> > {
> > gcc_assert (code == IFN_COND_ADD || code == IFN_COND_SUB
> > || code == IFN_COND_MUL || code == IFN_COND_AND
> > || code == IFN_COND_IOR || code == IFN_COND_XOR);
> > gcc_assert (op.num_ops == 4 && (op.ops[1] == op.ops[3]));
> > + cond_index = 0;
> > }
> >
> > bool masked_loop_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
> > @@ -8486,12 +8488,13 @@ vect_transform_reduction (loop_vec_info loop_vinfo,
> > vect_get_vec_defs (loop_vinfo, stmt_info, slp_node, ncopies,
> > single_defuse_cycle && reduc_index == 0
> > ? NULL_TREE : op.ops[0], &vec_oprnds0,
> > + cond_index == 0 ? truth_type_for (vectype_in) :
> > NULL_TREE,
> > single_defuse_cycle && reduc_index == 1
> > - ? NULL_TREE : op.ops[1], &vec_oprnds1,
> > + ? NULL_TREE : op.ops[1], &vec_oprnds1, NULL_TREE,
> > op.num_ops == 4
> > || (op.num_ops == 3
> > && !(single_defuse_cycle && reduc_index == 2))
> > - ? op.ops[2] : NULL_TREE, &vec_oprnds2);
> > + ? op.ops[2] : NULL_TREE, &vec_oprnds2, NULL_TREE);
>
> Ah, yes that's what I meant. I had something along those lines:
>
> + if (!cond_fn_p)
> + {
> + vect_get_vec_defs (loop_vinfo, stmt_info, slp_node, ncopies,
> + single_defuse_cycle && reduc_index == 0
> + ? NULL_TREE : op.ops[0], &vec_oprnds0,
> + single_defuse_cycle && reduc_index == 1
> + ? NULL_TREE : op.ops[1], &vec_oprnds1,
> + op.num_ops == 3
> + && !(single_defuse_cycle && reduc_index == 2)
> + ? op.ops[2] : NULL_TREE, &vec_oprnds2);
> + }
> + else
> + {
> + /* For a conditional operation, pass the truth type as vectype for the
> + mask. */
> + gcc_assert (single_defuse_cycle && reduc_index == 1);
> + vect_get_vec_defs (loop_vinfo, stmt_info, slp_node, ncopies,
> + op.ops[0], &vec_oprnds0,
> + truth_type_for (vectype_in),
> + NULL_TREE, &vec_oprnds1, NULL_TREE,
> + op.ops[2], &vec_oprnds2, NULL_TREE);
> + }
>
> Even though this has a bit of duplication now I prefer it slightly
> because of the. I'd hope, once fully tested (it's already running
> but I'm having connection problems so don't know about the results
> yet), this could go in independently of the pattern fix?
Yes, your version is also OK.
Thanks,
Richard.
>
> Regards
> Robin
>
