https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126516
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
I'll note the change will aggressively prefer SSE vectorization width,
irrespective of the cost of the operations leading up to the fold-left
reduction.
Also (but unlikely here), for slp_inst_kind_reduc_group we only
actually implement fold-left reductions for SLP_TREE_LANES == 1
unless there's no actual fold-left reduction to be performed
(VF == 1). Meaning, slp_inst_kind_reduc_group with VF == 1
is free, which isn't reflected in the change. But unlikely to be
the issue here.
So you might want to check whether the following fixes the regression:
diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
index dcfe4531f11..b2a96820ad3 100644
--- a/gcc/config/i386/i386.cc
+++ b/gcc/config/i386/i386.cc
@@ -27212,7 +27212,7 @@ ix86_vector_costs::finish_cost (const vector_costs
*scalar_costs)
{
unsigned int vf = vect_vf_for_cost (loop_vinfo);
for (auto inst : LOOP_VINFO_SLP_INSTANCES (loop_vinfo))
- if ((SLP_INSTANCE_KIND (inst) == slp_inst_kind_reduc_group
+ if (((SLP_INSTANCE_KIND (inst) == slp_inst_kind_reduc_group && vf != 1)
|| SLP_INSTANCE_KIND (inst) == slp_inst_kind_reduc_chain)
&& (vect_reduc_type (loop_vinfo, SLP_INSTANCE_TREE (inst))
== FOLD_LEFT_REDUCTION))
