There are some SVE testsuite fails when forcing SLP because costing
prevents VLA vectors from being used as we add permute cost for
the VEC_PERM nodes that are part of a SLP load-lanes node. The
permutes only exist for representational reasons and pessimize SLP
vs non-SLP so the following makes sure to cost them as zero.
Bootstrap and regtest running on x86_64-unknown-linux-gnu.
* tree-vect-slp.cc (vectorizable_slp_permutation_1): Return
zero for the permute nodes part of load-lanes.
---
gcc/tree-vect-slp.cc | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index 38f4c41b8db..1240dac3d62 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -10567,13 +10567,15 @@ vectorizable_slp_permutation_1 (vec_info *vinfo,
gimple_stmt_iterator *gsi,
/* Load-lanes permute. This permute only acts as a forwarder to
select the correct vector def of the load-lanes load which
has the permuted vectors in its vector defs like
- { v0, w0, r0, v1, w1, r1 ... } for a ld3. */
+ { v0, w0, r0, v1, w1, r1 ... } for a ld3. All costs are
+ accounted for in the costing for the actual load so we
+ return zero here. */
if (node->ldst_lanes)
{
gcc_assert (children.length () == 1);
if (!gsi)
/* This is a trivial op always supported. */
- return 1;
+ return 0;
slp_tree child = children[0];
unsigned vec_idx = (SLP_TREE_LANE_PERMUTATION (node)[0].second
/ SLP_TREE_LANES (node));
@@ -10583,7 +10585,7 @@ vectorizable_slp_permutation_1 (vec_info *vinfo,
gimple_stmt_iterator *gsi,
tree def = SLP_TREE_VEC_DEFS (child)[i * vec_num + vec_idx];
node->push_vec_def (def);
}
- return 1;
+ return 0;
}
/* Set REPEATING_P to true if the permutations are cylical wrt UNPACK_FACTOR
--
2.43.0
