The following properly guards the re-align (optimized) paths used
on old power CPUs for the added case of SLP splats from non-grouped
loads. Testcases are existing in dg-torture.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
PR tree-optimization/110614
* tree-vect-data-refs.cc (vect_supportable_dr_alignment):
SLP splats are not suitable for re-align ops.
---
gcc/tree-vect-data-refs.cc | 9 +++++----
1 file changed, 5 insertions(+), 4 deletions(-)
diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
index ab2af103cb4..9edc8989de9 100644
--- a/gcc/tree-vect-data-refs.cc
+++ b/gcc/tree-vect-data-refs.cc
@@ -6829,10 +6829,11 @@ vect_supportable_dr_alignment (vec_info *vinfo,
dr_vec_info *dr_info,
same alignment, instead it depends on the SLP group size. */
if (loop_vinfo
&& STMT_SLP_TYPE (stmt_info)
- && !multiple_p (LOOP_VINFO_VECT_FACTOR (loop_vinfo)
- * (DR_GROUP_SIZE
- (DR_GROUP_FIRST_ELEMENT (stmt_info))),
- TYPE_VECTOR_SUBPARTS (vectype)))
+ && (!STMT_VINFO_GROUPED_ACCESS (stmt_info)
+ || !multiple_p (LOOP_VINFO_VECT_FACTOR (loop_vinfo)
+ * (DR_GROUP_SIZE
+ (DR_GROUP_FIRST_ELEMENT (stmt_info))),
+ TYPE_VECTOR_SUBPARTS (vectype))))
;
else if (!loop_vinfo
|| (nested_in_vect_loop
--
2.35.3
