[PATCH] 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.

Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.

	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.
---
 gcc/testsuite/gcc.dg/vect/bb-slp-pr126926.c | 37 +++++++++++++++++++++
 gcc/tree-vect-slp.cc                        | 26 ++++++---------
 2 files changed, 48 insertions(+), 15 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/vect/bb-slp-pr126926.c

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 00000000000..84ebf5cc93d
--- /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 f40ddd4b2b7..5e729e478ce 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -12551,19 +12551,17 @@ 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
 	{
@@ -12578,12 +12576,15 @@ 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)
     {
@@ -12686,11 +12687,6 @@ 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
-- 
2.51.0