Hi,
in VMAT_STRIDED_SLP we're likely to select a different vectype with
fewer elements for vector construction. After loading it is
re-interpreted as the proper vectype.
When checking costs we use the original vectype with more elements
leading to wrong costing in case vector construction is dependent
on the number of elements.
This patch makes a temporary copy of stmt_info and slp_node, changes
their vectype and passes them to record_stmt_cost in case we chose
a different load/construction vectype.
Bootstrapped and regtested on x86, aarch64 and power10.
Regtested on rv64gcv.
Regards
Robin
PR target/118019
gcc/ChangeLog:
* tree-vect-stmts.cc (vectorizable_load): Use construction/load
vectype for costing.
---
gcc/tree-vect-stmts.cc | 23 ++++++++++++++++++++---
1 file changed, 20 insertions(+), 3 deletions(-)
diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index be1139a423c..6ac1e97c4c1 100644
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -10810,9 +10810,26 @@ vectorizable_load (vec_info *vinfo,
if (nloads > 1)
{
if (costing_p)
- inside_cost += record_stmt_cost (cost_vec, 1, vec_construct,
- stmt_info, slp_node, 0,
- vect_body);
+ {
+ if (lvectype != vectype)
+ {
+ /* If we chose a different vectype for vector
+ construction make sure to use it for costing. */
+ stmt_vec_info stmt_info_copy = stmt_info;
+ stmt_info_copy->vectype = lvectype;
+ slp_tree slp_node_copy = slp_node;
+ slp_node_copy->vectype = lvectype;
+ inside_cost
+ += record_stmt_cost (cost_vec, 1, vec_construct,
+ stmt_info_copy, slp_node_copy,
+ 0, vect_body);
+ }
+
+ else
+ inside_cost += record_stmt_cost (cost_vec, 1, vec_construct,
+ stmt_info, slp_node, 0,
+ vect_body);
+ }
else
{
tree vec_inv = build_constructor (lvectype, v);
--
2.47.1
