https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110742
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rsandifo at gcc dot gnu.org
--- Comment #12 from Richard Biener <rguenth at gcc dot gnu.org> ---
So we have those external def nodes for existing vectors in the IL. Obviously
we cannot change the layout of those - where can we specify that? I
see is_compatible_layout or change_layout_cost. But start_choosing_layouts
already assigns layout 0 to it where I thought this would effectively
prevent changes?
Oddly in forward_pass we first have partition.layout == 0 && layout_i == 0
but still do stuff.
$38 = 0x3d8f690 = {{node_begin = 0, node_end = 1, layout = 0, in_degree = 0,
out_degree = 1}, {
node_begin = 1, node_end = 2, layout = 1, in_degree = 1, out_degree = 0},
{node_begin = 2,
node_end = 3, layout = 2, in_degree = 0, out_degree = 0}}
later we ask if changing layout from 0 to 1 is OK.
I suppose the bug is in get_result_with_layout itself which assumes
changing vect_constant_def/vect_external_def is fine and thus pushes
the edge permute to the node?
I wonder if I fix that say with the following, if I also need to adjust
costs somewhere?
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index 693621ca990..1d79c77e8ce 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -5198,7 +5198,10 @@ vect_optimize_slp_pass::get_result_with_layout (slp_tree
node,
return result;
if (SLP_TREE_DEF_TYPE (node) == vect_constant_def
- || SLP_TREE_DEF_TYPE (node) == vect_external_def)
+ || (SLP_TREE_DEF_TYPE (node) == vect_external_def
+ && (to_layout_i == 0
+ /* We can't permute vector defs. */
+ || SLP_TREE_VEC_DEFS (node).is_empty ())))
{
/* If the vector is uniform or unchanged, there's nothing to do. */
if (to_layout_i == 0 || vect_slp_tree_uniform_p (node))
I'll also note that we do
/* Handle externals and constants optimistically throughout.
But treat existing vectors as fixed since we do not handle
permuting them. */
unsigned int node_i = m_partitioned_nodes[rpo_i];
auto &vertex = m_vertices[node_i];
if ((SLP_TREE_DEF_TYPE (vertex.node) == vect_external_def
&& !SLP_TREE_VEC_DEFS (vertex.node).exists ())
|| SLP_TREE_DEF_TYPE (vertex.node) == vect_constant_def)
vertex.partition = -1;
in create_partitions ().
I'm testing the above change now.
