When materializing on a VEC_PERM node we have to permute the
incoming vectors, not the outgoing one.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
2021-01-05 Richard Biener <rguent...@suse.de>
PR tree-optimization/98516
* tree-vect-slp.c (vect_optimize_slp): Permute the incoming
lanes when materializing on a VEC_PERM node.
(vectorizable_slp_permutation): Dump the permute properly.
* gcc.dg/vect/bb-slp-pr98516-1.c: New testcase.
* gcc.dg/vect/bb-slp-pr98516-2.c: Likewise.
---
gcc/testsuite/gcc.dg/vect/bb-slp-pr98516-1.c | 26 ++++++++++++++
gcc/testsuite/gcc.dg/vect/bb-slp-pr98516-2.c | 36 ++++++++++++++++++++
gcc/tree-vect-slp.c | 13 ++++---
3 files changed, 70 insertions(+), 5 deletions(-)
create mode 100644 gcc/testsuite/gcc.dg/vect/bb-slp-pr98516-1.c
create mode 100644 gcc/testsuite/gcc.dg/vect/bb-slp-pr98516-2.c
diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-pr98516-1.c
b/gcc/testsuite/gcc.dg/vect/bb-slp-pr98516-1.c
new file mode 100644
index 00000000000..c4c244c6f8a
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/bb-slp-pr98516-1.c
@@ -0,0 +1,26 @@
+/* { dg-do run } */
+
+double a[4], b[2];
+
+void __attribute__((noipa))
+foo ()
+{
+ double a0 = a[0];
+ double a1 = a[1];
+ double a2 = a[2];
+ double a3 = a[3];
+ b[0] = a1 - a3;
+ b[1] = a0 + a2;
+}
+
+int main()
+{
+ a[0] = 1.;
+ a[1] = 2.;
+ a[2] = 3.;
+ a[3] = 4.;
+ foo ();
+ if (b[0] != -2 || b[1] != 4)
+ __builtin_abort ();
+ return 0;
+}
diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-pr98516-2.c
b/gcc/testsuite/gcc.dg/vect/bb-slp-pr98516-2.c
new file mode 100644
index 00000000000..f1a9341e224
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/bb-slp-pr98516-2.c
@@ -0,0 +1,36 @@
+/* { dg-do run } */
+
+float a[8], b[4];
+
+void __attribute__((noipa))
+foo ()
+{
+ float a0 = a[0];
+ float a1 = a[1];
+ float a2 = a[2];
+ float a3 = a[3];
+ float a4 = a[4];
+ float a5 = a[5];
+ float a6 = a[6];
+ float a7 = a[7];
+ b[0] = a1 - a5;
+ b[1] = a0 + a4;
+ b[2] = a3 - a7;
+ b[3] = a2 + a6;
+}
+
+int main()
+{
+ a[0] = 1.;
+ a[1] = 2.;
+ a[2] = 3.;
+ a[3] = 4.;
+ a[4] = 5.;
+ a[5] = 6.;
+ a[6] = 7.;
+ a[7] = 8.;
+ foo ();
+ if (b[0] != -4 || b[1] != 6 || b[2] != -4 || b[3] != 10)
+ __builtin_abort ();
+ return 0;
+}
diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c
index 49cb635ee92..c9da8457e5e 100644
--- a/gcc/tree-vect-slp.c
+++ b/gcc/tree-vect-slp.c
@@ -3094,15 +3094,18 @@ vect_optimize_slp (vec_info *vinfo)
;
else if (SLP_TREE_LANE_PERMUTATION (node).exists ())
{
- /* If the node if already a permute node we just need to apply
- the permutation to the permute node itself. */
+ /* If the node is already a permute node we can apply
+ the permutation to the lane selection, effectively
+ materializing it on the incoming vectors. */
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"simplifying permute node %p\n",
node);
- vect_slp_permute (perms[perm], SLP_TREE_LANE_PERMUTATION (node),
- true);
+ for (unsigned k = 0;
+ k < SLP_TREE_LANE_PERMUTATION (node).length (); ++k)
+ SLP_TREE_LANE_PERMUTATION (node)[k].second
+ = perms[perm][SLP_TREE_LANE_PERMUTATION (node)[k].second];
}
else
{
@@ -5554,7 +5557,7 @@ vectorizable_slp_permutation (vec_info *vinfo,
gimple_stmt_iterator *gsi,
dump_printf (MSG_NOTE, ",");
dump_printf (MSG_NOTE, " vops%u[%u][%u]",
vperm[i].first.first, vperm[i].first.second,
- vperm[i].first.second);
+ vperm[i].second);
}
dump_printf (MSG_NOTE, "\n");
}
--
2.26.2
