[gcc r16-9556] tree-optimization/126926 - back-to-back SLP instance scheduling

https://gcc.gnu.org/g:df724d7f92e60cc8c21cda11d0958f4ec69504ee

commit r16-9556-gdf724d7f92e60cc8c21cda11d0958f4ec69504ee
Author: Richard Biener <[email protected]>
Date:   Tue Aug 18 11:02:12 2026 +0200

    tree-optimization/126926 - back-to-back SLP instance scheduling
    
    When we have two SLP instances interacting at the
    c_13 definition point of a vector which is both a vector CTOR root
    for one instance and the leaf (existing vector) of a reduction instance,
    we transform the CTOR instance first replacing the definition
    in vectorize_slp_instance_root_stmt.  That disrupts the SLP scheduling
    process of the 2nd as the replacement now appears to be outside of the
    SLP region.  Avoid this by not replacing the vector constructing stmt but
    only its RHS.
    
            PR tree-optimization/126926
            * tree-vect-slp.cc (vectorize_slp_instance_root_stmt): Only
            replace the RHS of the vector constructing scalar stmt.
    
            * gcc.dg/vect/bb-slp-pr126926.c: New testcase.
    
    (cherry picked from commit a9c22e6d049bf10e44b6a76384a1eb00991c3eb8)

Diff:
---
 gcc/testsuite/gcc.dg/vect/bb-slp-pr126926.c | 37 +++++++++++++++++++++++++++++
 gcc/tree-vect-slp.cc                        | 26 +++++++++-----------
 2 files changed, 48 insertions(+), 15 deletions(-)

diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-pr126926.c b/gcc/testsuite/gcc.dg/vect/bb-slp-pr126926.c
new file mode 100644
index 000000000000..84ebf5cc93d8
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/bb-slp-pr126926.c
@@ -0,0 +1,37 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-fgimple" } */
+
+float a;
+
+void __GIMPLE (ssa,guessed_local(1073741824),startwith("slp"))
+b ()
+{
+  float [[gnu::vector_size(8)]] c;
+  float _1;
+  float _2;
+  float _3;
+  float _4;
+  float _5;
+  float _6;
+  float _7;
+  float _8;
+  float _9;
+
+  __BB(2,guessed_local(1073741824)):
+  _1 = a;
+  _2 = _1 + _Literal (float) 0.0;
+  c_12 = __BIT_INSERT (c_11(D), _2, 32u);
+  _3 = a;
+  _4 = _3 + _Literal (float) 0.0;
+  c_13 = __BIT_INSERT (c_12, _4, 0u);
+  _5 = __BIT_FIELD_REF <float> (c_13, 32u, 32u);
+  _6 = _5 * _Literal (float) 0.0;
+  _7 = __BIT_FIELD_REF <float> (c_13, 32u, 32u);
+  _8 = _7 * _Literal (float) 2.0e+0;
+  _9 = _6 + _8;
+  a = _9;
+  return;
+
+}
+
+
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index 48bfd87f8961..94b6aad94008 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -12198,19 +12198,17 @@ vect_remove_slp_scalar_calls (vec_info *vinfo, slp_tree node)
 void
 vectorize_slp_instance_root_stmt (vec_info *vinfo, slp_tree node, slp_instance instance)
 {
-  gassign *rstmt = NULL;
-
   if (instance->kind == slp_inst_kind_ctor)
     {
+      tree new_def;
       if (SLP_TREE_VEC_DEFS (node).length () == 1)
 	{
-	  tree vect_lhs = SLP_TREE_VEC_DEFS (node)[0];
+	  new_def = SLP_TREE_VEC_DEFS (node)[0];
 	  tree root_lhs = gimple_get_lhs (instance->root_stmts[0]->stmt);
 	  if (!useless_type_conversion_p (TREE_TYPE (root_lhs),
-					  TREE_TYPE (vect_lhs)))
-	    vect_lhs = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (root_lhs),
-			       vect_lhs);
-	  rstmt = gimple_build_assign (root_lhs, vect_lhs);
+					  TREE_TYPE (new_def)))
+	    new_def = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (root_lhs),
+			       new_def);
 	}
       else
 	{
@@ -12225,12 +12223,15 @@ vectorize_slp_instance_root_stmt (vec_info *vinfo, slp_tree node, slp_instance i
 	     do not match.  */
 	  FOR_EACH_VEC_ELT (SLP_TREE_VEC_DEFS (node), j, child_def)
 	    CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, child_def);
-	  tree lhs = gimple_get_lhs (instance->root_stmts[0]->stmt);
 	  tree rtype
 	    = TREE_TYPE (gimple_assign_rhs1 (instance->root_stmts[0]->stmt));
-	  tree r_constructor = build_constructor (rtype, v);
-	  rstmt = gimple_build_assign (lhs, r_constructor);
+	  new_def = build_constructor (rtype, v);
 	}
+
+      gimple_stmt_iterator rgsi = gsi_for_stmt (instance->root_stmts[0]->stmt);
+      gimple_assign_set_rhs_from_tree (&rgsi, new_def);
+      update_stmt (gsi_stmt (rgsi));
+      return;
     }
   else if (instance->kind == slp_inst_kind_bb_reduc)
     {
@@ -12316,11 +12317,6 @@ vectorize_slp_instance_root_stmt (vec_info *vinfo, slp_tree node, slp_instance i
     }
   else
     gcc_unreachable ();
-
-  gcc_assert (rstmt);
-
-  gimple_stmt_iterator rgsi = gsi_for_stmt (instance->root_stmts[0]->stmt);
-  gsi_replace (&rgsi, rstmt, true);
 }
 
 struct slp_scc_info