re PR tree-optimization/65930 (Reduction with sign-change not handled)

2019-10-30  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/65930
	* tree-vect-loop.c (vect_is_simple_reduction): For reduction
	chains also allow a leading and trailing conversion.
	* tree-vect-slp.c (vect_get_and_check_slp_defs): Handle
	intermediate reduction chains.
	(vect_analyze_slp_instance): Likewise.  Build a SLP
	node for a trailing conversion manually.

	* gcc.dg/vect/pr65930-2.c: New testcase.

From-SVN: r277603

5 files changed, 91 insertions, 5 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 72d6ec7..618197d 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,13 @@
+2019-10-30  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/65930
+	* tree-vect-loop.c (vect_is_simple_reduction): For reduction
+	chains also allow a leading and trailing conversion.
+	* tree-vect-slp.c (vect_get_and_check_slp_defs): Handle
+	intermediate reduction chains.
+	(vect_analyze_slp_instance): Likewise.  Build a SLP
+	node for a trailing conversion manually.
+
 2019-10-30  Martin Liska  <mliska@suse.cz>
 
 	* cgraph.c (cgraph_node::local_info): Transform to ...
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 519952f..23caef0 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2019-10-30  Richard Biener  <rguenther@suse.de>
+
+	PR tree-optimization/65930
+	* gcc.dg/vect/pr65930-2.c: New testcase.
+
 2019-10-30  Mark Eggleston <mark.eggleston@codethink.com>
 
 	* gfortran.dg/no_overwrite_recursive_1.f90: New test.
diff --git a/gcc/testsuite/gcc.dg/vect/pr65930-2.c b/gcc/testsuite/gcc.dg/vect/pr65930-2.c
new file mode 100644
index 0000000..9cfb9b1
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr65930-2.c
@@ -0,0 +1,28 @@
+/* { dg-require-effective-target vect_int } */
+
+#include "tree-vect.h"
+
+int __attribute__((noipa))
+bar (unsigned int *x, int n)
+{
+  int sum = 4;
+  x = __builtin_assume_aligned (x, __BIGGEST_ALIGNMENT__);
+  for (int i = 0; i < n; ++i)
+    sum += x[i*4+0]+ x[i*4 + 1] + x[i*4 + 2] + x[i*4 + 3];
+  return sum;
+}
+
+int
+main ()
+{
+  static int a[16] __attribute__((aligned(__BIGGEST_ALIGNMENT__)))
+    = { 1, 3, 5, 8, 9, 10, 17, 18, 23, 29, 30, 55, 42, 2, 3, 1 };
+  check_vect ();
+  if (bar (a, 4) != 260)
+    abort ();
+  return 0;
+}
+
+/* { dg-final { scan-tree-dump "vectorized 1 loops" "vect" } } */
+/* { dg-final { scan-tree-dump "Loop contains only SLP stmts" "vect" } } */
+/* { dg-final { scan-tree-dump "vectorizing stmts using SLP" "vect" } } */
diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
index 9b7d248..fa873e9 100644
--- a/gcc/tree-vect-loop.c
+++ b/gcc/tree-vect-loop.c
@@ -3005,14 +3005,21 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
       for (i = path.length () - 1; i >= 1; --i)
 	{
 	  gimple *stmt = USE_STMT (path[i].second);
-	  if (gimple_assign_rhs_code (stmt) != code
+	  stmt_vec_info stmt_info = loop_info->lookup_stmt (stmt);
+	  STMT_VINFO_REDUC_IDX (stmt_info)
+	    = path[i].second->use - gimple_assign_rhs1_ptr (stmt);
+	  enum tree_code stmt_code = gimple_assign_rhs_code (stmt);
+	  bool leading_conversion = (CONVERT_EXPR_CODE_P (stmt_code)
+				     && (i == 1 || i == path.length () - 1));
+	  if ((stmt_code != code && !leading_conversion)
 	      /* We can only handle the final value in epilogue
 		 generation for reduction chains.  */
 	      || (i != 1 && !has_single_use (gimple_assign_lhs (stmt))))
 	    is_slp_reduc = false;
-	  stmt_vec_info stmt_info = loop_info->lookup_stmt (stmt);
-	  STMT_VINFO_REDUC_IDX (stmt_info)
-	    = path[i].second->use - gimple_assign_rhs1_ptr (stmt);
+	  /* For reduction chains we support a trailing/leading
+	     conversions.  We do not store those in the actual chain.  */
+	  if (leading_conversion)
+	    continue;
 	  reduc_chain.safe_push (stmt_info);
 	}
       if (is_slp_reduc && reduc_chain.length () > 1)
diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c
index 9d65471..4b1a231 100644
--- a/gcc/tree-vect-slp.c
+++ b/gcc/tree-vect-slp.c
@@ -419,6 +419,13 @@ again:
 
       if (first)
 	{
+	  /* For the swapping logic below force vect_reduction_def
+	     for the reduction op in a SLP reduction group.  */
+	  if (!STMT_VINFO_DATA_REF (stmt_info)
+	      && REDUC_GROUP_FIRST_ELEMENT (stmt_info)
+	      && (int)i == STMT_VINFO_REDUC_IDX (stmt_info)
+	      && def_stmt_info)
+	    dt = vect_reduction_def;
 	  oprnd_info->first_dt = dt;
 	  oprnd_info->first_op_type = TREE_TYPE (oprnd);
 	}
@@ -2041,7 +2048,8 @@ vect_analyze_slp_instance (vec_info *vinfo,
       /* Mark the first element of the reduction chain as reduction to properly
 	 transform the node.  In the reduction analysis phase only the last
 	 element of the chain is marked as reduction.  */
-      STMT_VINFO_DEF_TYPE (stmt_info) = vect_reduction_def;
+      STMT_VINFO_DEF_TYPE (stmt_info)
+	= STMT_VINFO_DEF_TYPE (scalar_stmts.last ());
       STMT_VINFO_REDUC_DEF (vect_orig_stmt (stmt_info))
 	= STMT_VINFO_REDUC_DEF (vect_orig_stmt (scalar_stmts.last ()));
     }
@@ -2071,6 +2079,34 @@ vect_analyze_slp_instance (vec_info *vinfo,
   delete bst_map;
   if (node != NULL)
     {
+      /* If this is a reduction chain with a conversion in front
+         amend the SLP tree with a node for that.  */
+      if (!dr
+	  && REDUC_GROUP_FIRST_ELEMENT (stmt_info)
+	  && STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def)
+	{
+	  /* Get at the conversion stmt - we know it's the single use
+	     of the last stmt of the reduction chain.  */
+	  gimple *tem = vect_orig_stmt (scalar_stmts[group_size - 1])->stmt;
+	  use_operand_p use_p;
+	  gimple *use_stmt;
+	  bool r = single_imm_use (gimple_assign_lhs (tem), &use_p, &use_stmt);
+	  gcc_assert (r);
+	  next_info = vinfo->lookup_stmt (use_stmt);
+	  next_info = vect_stmt_to_vectorize (next_info);
+	  scalar_stmts = vNULL;
+	  scalar_stmts.create (group_size);
+	  for (unsigned i = 0; i < group_size; ++i)
+	    scalar_stmts.quick_push (next_info);
+	  slp_tree conv = vect_create_new_slp_node (scalar_stmts);
+	  SLP_TREE_CHILDREN (conv).quick_push (node);
+	  node = conv;
+	  /* We also have to fake this conversion stmt as SLP reduction group
+	     so we don't have to mess with too much code elsewhere.  */
+	  REDUC_GROUP_FIRST_ELEMENT (next_info) = next_info;
+	  REDUC_GROUP_NEXT_ELEMENT (next_info) = NULL;
+	}
+
       /* Calculate the unrolling factor based on the smallest type.  */
       poly_uint64 unrolling_factor
 	= calculate_unrolling_factor (max_nunits, group_size);