[gcc r13-2492] tree-optimization/106841 - gather and hybrid SLP

[Richard Biener <rguenth@gcc.gnu.org>]

https://gcc.gnu.org/g:e33e61d417eb5e981bb7d709f8681a2f55ed518a

commit r13-2492-ge33e61d417eb5e981bb7d709f8681a2f55ed518a
Author: Richard Biener <rguenther@suse.de>
Date:   Tue Sep 6 10:08:44 2022 +0200

    tree-optimization/106841 - gather and hybrid SLP
    
    Hybrid SLP detection currently fails to consider a not direct
    offset operand of a scatter/gather operation.  The following fixes
    this.
    
            PR tree-optimization/106841
            * tree-vect-slp.cc (vect_detect_hybrid_slp): Also process
            scatter/gather offset.
    
            * g++.dg/vect/pr106841.cc: New testcase.

Diff:
---
 gcc/testsuite/g++.dg/vect/pr106841.cc | 52 +++++++++++++++++++++++++++++++++++
 gcc/tree-vect-slp.cc                  |  9 ++++++
 2 files changed, 61 insertions(+)

diff --git a/gcc/testsuite/g++.dg/vect/pr106841.cc b/gcc/testsuite/g++.dg/vect/pr106841.cc
new file mode 100644
index 00000000000..7458bc15a25
--- /dev/null
+++ b/gcc/testsuite/g++.dg/vect/pr106841.cc
@@ -0,0 +1,52 @@
+// { dg-do compile }
+// { dg-additional-options "-O3 -ffast-math" }
+// { dg-additional-options "-march=bdver2" { target x86_64-*-* } }
+
+struct R3 {
+  double z;
+  R3(R3 A, R3 B) : z(B.z - A.z) {}
+  double norme() { return z; }
+};
+struct TBoundaryEdge {
+  int *vertices[2];
+  int &operator[](int i) { return *vertices[i]; }
+};
+struct Mesh {
+  int vertices;
+  TBoundaryEdge *bedges;
+  int operator()(int &vv) { return &vv - &vertices; }
+  TBoundaryEdge be(int i) { return bedges[i]; }
+};
+template <typename Data> struct GenericElement {
+  typedef typename Data::V Vertex;
+  static const int nv = Data::NbOfVertices;
+  Vertex *vertices[nv];
+  double mes;
+  void set(int *iv, Vertex *v0) {
+    for (int i = 0; i < nv; ++i)
+      vertices[i] = v0 + iv[i];
+    mes = Data::mesure(vertices);
+  }
+};
+struct DataSeg3 {
+  static const int NbOfVertices = 2;
+  typedef R3 V;
+  static double mesure(V *pv[]) { return R3(*pv[0], *pv[1]).norme(); }
+};
+struct MeshS {
+  MeshS();
+};
+template <class> struct Movemesh_Op { void foo(Mesh, DataSeg3::V *) const; };
+template <> void Movemesh_Op<int>::foo(Mesh pTh, DataSeg3::V *v0) const {
+  GenericElement<DataSeg3> *bS = new GenericElement<DataSeg3>[8];
+  for (int ibe = 0; ibe < 8; ibe++) {
+    TBoundaryEdge K(pTh.be(ibe));
+    int iv[2];
+    for (int i = 0; i < 2; i++) {
+      int &__trans_tmp_2 = K[i];
+      iv[i] = pTh(__trans_tmp_2);
+    }
+    bS[ibe].set(iv, v0);
+  }
+  MeshS T_Th;
+}
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index b10f69da133..3fa2dc96dda 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -5799,6 +5799,15 @@ vect_detect_hybrid_slp (loop_vec_info loop_vinfo)
 	 to use walk_gimple_op.  */
       wi.is_lhs = 0;
       walk_gimple_op (stmt_info->stmt, vect_detect_hybrid_slp, &wi);
+      /* For gather/scatter make sure to walk the offset operand, that
+	 can be a scaling and conversion away.  */
+      gather_scatter_info gs_info;
+      if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)
+	  && vect_check_gather_scatter (stmt_info, loop_vinfo, &gs_info))
+	{
+	  int dummy;
+	  vect_detect_hybrid_slp (&gs_info.offset, &dummy, &wi);
+	}
     }
 }