]> git.ipfire.org Git - thirdparty/gcc.git/commitdiff
2019-10-16 Richard Biener <rguenther@suse.de>
authorrguenth <rguenth@138bc75d-0d04-0410-961f-82ee72b054a4>
Wed, 16 Oct 2019 14:21:06 +0000 (14:21 +0000)
committerrguenth <rguenth@138bc75d-0d04-0410-961f-82ee72b054a4>
Wed, 16 Oct 2019 14:21:06 +0000 (14:21 +0000)
* tree-vect-loop.c (vect_valid_reduction_input_p): Remove.
(vect_is_simple_reduction): Delay checking to
vectorizable_reduction and relax the checking.
(vectorizable_reduction): Check we have a simple use.  Check
for bogus condition reductions.
* tree-vect-stmts.c (vect_transform_stmt): Make sure we
are looking at the last stmt in a pattern sequence when
filling in backedge PHI values.

* gcc.dg/vect/vect-cond-reduc-3.c: New testcase.
* gcc.dg/vect/vect-cond-reduc-4.c: Likewise.

git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@277067 138bc75d-0d04-0410-961f-82ee72b054a4

gcc/ChangeLog
gcc/testsuite/ChangeLog
gcc/testsuite/gcc.dg/vect/vect-cond-reduc-3.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/vect/vect-cond-reduc-4.c [new file with mode: 0644]
gcc/tree-vect-loop.c
gcc/tree-vect-stmts.c

index da7bf1fddbfdd5e88bd7ab4dc13f3d3b2db84f35..a6047ec1008cafe9f9addcad7cd0f48e8bf0f86f 100644 (file)
@@ -1,3 +1,14 @@
+2019-10-16  Richard Biener  <rguenther@suse.de>
+
+       * tree-vect-loop.c (vect_valid_reduction_input_p): Remove.
+       (vect_is_simple_reduction): Delay checking to
+       vectorizable_reduction and relax the checking.
+       (vectorizable_reduction): Check we have a simple use.  Check
+       for bogus condition reductions.
+       * tree-vect-stmts.c (vect_transform_stmt): Make sure we
+       are looking at the last stmt in a pattern sequence when
+       filling in backedge PHI values.
+
 2019-10-16  Peter Bergner <bergner@linux.ibm.com>
            Jiufu Guo  <guojiufu@linux.ibm.com>
 
index 93ca734f8a90483da0174c83f4c30e330a1606b6..cd43ef621066785ec33b5b5c4052390b780b2f8a 100644 (file)
@@ -1,3 +1,8 @@
+2019-10-16  Richard Biener  <rguenther@suse.de>
+
+       * gcc.dg/vect/vect-cond-reduc-3.c: New testcase.
+       * gcc.dg/vect/vect-cond-reduc-4.c: Likewise.
+
 2019-10-16  Peter Bergner <bergner@linux.ibm.com>
            Jiufu Guo  <guojiufu@linux.ibm.com>
 
diff --git a/gcc/testsuite/gcc.dg/vect/vect-cond-reduc-3.c b/gcc/testsuite/gcc.dg/vect/vect-cond-reduc-3.c
new file mode 100644 (file)
index 0000000..a5b3849
--- /dev/null
@@ -0,0 +1,45 @@
+/* { dg-require-effective-target vect_condition } */
+/* { dg-require-effective-target vect_float } */
+
+#include "tree-vect.h"
+
+extern void abort (void) __attribute__ ((noreturn));
+
+#define N 27
+
+/* Condition reduction with different types.  */
+
+int
+condition_reduction (float *a, float min_v)
+{
+  int last = 0;
+
+  for (int i = 0; i < N; i++)
+    if (a[i] < min_v)
+      last = i;
+
+  return last;
+}
+
+int
+main (void)
+{
+  float a[N] = {
+  11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
+  1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
+  21, 22, 23, 24, 25, 26, 27
+  };
+
+  check_vect ();
+
+  int ret = condition_reduction (a, 10);
+  if (ret != 18)
+    abort ();
+
+  return 0;
+}
+
+/* { dg-final { scan-tree-dump-times "LOOP VECTORIZED" 2 "vect" } } */
+/* { dg-final { scan-tree-dump-times "optimizing condition reduction with FOLD_EXTRACT_LAST" 4 "vect" { target vect_fold_extract_last } } } */
+/* { dg-final { scan-tree-dump-times "condition expression based on integer induction." 2 "vect" { target { ! vect_fold_extract_last } } } } */
+
diff --git a/gcc/testsuite/gcc.dg/vect/vect-cond-reduc-4.c b/gcc/testsuite/gcc.dg/vect/vect-cond-reduc-4.c
new file mode 100644 (file)
index 0000000..6b6d17f
--- /dev/null
@@ -0,0 +1,45 @@
+/* { dg-require-effective-target vect_condition } */
+/* { dg-require-effective-target vect_double } */
+
+#include "tree-vect.h"
+
+extern void abort (void) __attribute__ ((noreturn));
+
+#define N 27
+
+/* Condition reduction with different types.  */
+
+int
+condition_reduction (double *a, double min_v)
+{
+  int last = 0;
+
+  for (int i = 0; i < N; i++)
+    if (a[i] < min_v)
+      last = i;
+
+  return last;
+}
+
+int
+main (void)
+{
+  double a[N] = {
+  11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
+  1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
+  21, 22, 23, 24, 25, 26, 27
+  };
+
+  check_vect ();
+
+  int ret = condition_reduction (a, 10);
+  if (ret != 18)
+    abort ();
+
+  return 0;
+}
+
+/* { dg-final { scan-tree-dump-times "LOOP VECTORIZED" 2 "vect" } } */
+/* { dg-final { scan-tree-dump-times "optimizing condition reduction with FOLD_EXTRACT_LAST" 4 "vect" { target vect_fold_extract_last } } } */
+/* { dg-final { scan-tree-dump-times "condition expression based on integer induction." 2 "vect" { target { ! vect_fold_extract_last } } } } */
+
index 455251070d081d11f5301f7178dc2d62f2f2bb0e..0530d6643b46b8af306f1eae001682b189c46055 100644 (file)
@@ -2532,21 +2532,6 @@ report_vect_op (dump_flags_t msg_type, gimple *stmt, const char *msg)
   dump_printf_loc (msg_type, vect_location, "%s%G", msg, stmt);
 }
 
-/* DEF_STMT_INFO occurs in a loop that contains a potential reduction
-   operation.  Return true if the results of DEF_STMT_INFO are something
-   that can be accumulated by such a reduction.  */
-
-static bool
-vect_valid_reduction_input_p (stmt_vec_info def_stmt_info)
-{
-  return (is_gimple_assign (def_stmt_info->stmt)
-         || is_gimple_call (def_stmt_info->stmt)
-         || STMT_VINFO_DEF_TYPE (def_stmt_info) == vect_induction_def
-         || (gimple_code (def_stmt_info->stmt) == GIMPLE_PHI
-             && STMT_VINFO_DEF_TYPE (def_stmt_info) == vect_internal_def
-             && !is_loop_header_bb_p (gimple_bb (def_stmt_info->stmt))));
-}
-
 /* Return true if we need an in-order reduction for operation CODE
    on type TYPE.  NEED_WRAPPING_INTEGRAL_OVERFLOW is true if integer
    overflow must wrap.  */
@@ -2748,13 +2733,7 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
                          bool *double_reduc)
 {
   gphi *phi = as_a <gphi *> (phi_info->stmt);
-  class loop *loop = (gimple_bb (phi))->loop_father;
-  class loop *vect_loop = LOOP_VINFO_LOOP (loop_info);
-  bool nested_in_vect_loop = flow_loop_nested_p (vect_loop, loop);
   gimple *phi_use_stmt = NULL;
-  enum tree_code orig_code, code;
-  tree op1, op2, op3 = NULL_TREE, op4 = NULL_TREE;
-  tree type;
   imm_use_iterator imm_iter;
   use_operand_p use_p;
 
@@ -2768,6 +2747,7 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
      can be constant.  See PR60382.  */
   if (has_zero_uses (phi_name))
     return NULL;
+  class loop *loop = (gimple_bb (phi))->loop_father;
   unsigned nphi_def_loop_uses = 0;
   FOR_EACH_IMM_USE_FAST (use_p, imm_iter, phi_name)
     {
@@ -2802,6 +2782,8 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
       || !flow_bb_inside_loop_p (loop, gimple_bb (def_stmt_info->stmt)))
     return NULL;
 
+  bool nested_in_vect_loop
+    = flow_loop_nested_p (LOOP_VINFO_LOOP (loop_info), loop);
   unsigned nlatch_def_loop_uses = 0;
   auto_vec<gphi *, 3> lcphis;
   bool inner_loop_of_double_reduc = false;
@@ -2849,8 +2831,7 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
      defined in the inner loop.  */
   if (gphi *def_stmt = dyn_cast <gphi *> (def_stmt_info->stmt))
     {
-      op1 = PHI_ARG_DEF (def_stmt, 0);
-
+      tree op1 = PHI_ARG_DEF (def_stmt, 0);
       if (gimple_phi_num_args (def_stmt) != 1
           || TREE_CODE (op1) != SSA_NAME)
         {
@@ -2890,7 +2871,7 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
                         def_stmt_info->stmt);
       return NULL;
     }
-  code = orig_code = gimple_assign_rhs_code (def_stmt);
+  enum tree_code code = gimple_assign_rhs_code (def_stmt);
 
   /* We can handle "res -= x[i]", which is non-associative by
      simply rewriting this into "res += -x[i]".  Avoid changing
@@ -2899,26 +2880,11 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
   if (code == MINUS_EXPR && gimple_assign_rhs2 (def_stmt) != phi_name)
     code = PLUS_EXPR;
 
+  tree op1, op2;
   if (code == COND_EXPR)
     {
       if (! nested_in_vect_loop)
        STMT_VINFO_REDUC_TYPE (phi_info) = COND_REDUCTION;
-
-      op3 = gimple_assign_rhs1 (def_stmt);
-      if (COMPARISON_CLASS_P (op3))
-        {
-          op4 = TREE_OPERAND (op3, 1);
-          op3 = TREE_OPERAND (op3, 0);
-        }
-      if (op3 == phi_name || op4 == phi_name)
-       {
-         if (dump_enabled_p ())
-           report_vect_op (MSG_MISSED_OPTIMIZATION, def_stmt,
-                           "reduction: condition depends on previous"
-                           " iteration: ");
-         return NULL;
-       }
-
       op1 = gimple_assign_rhs2 (def_stmt);
       op2 = gimple_assign_rhs3 (def_stmt);
     }
@@ -2951,33 +2917,6 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
       return NULL;
     }
 
-  type = TREE_TYPE (gimple_assign_lhs (def_stmt));
-  if ((TREE_CODE (op1) == SSA_NAME
-       && !types_compatible_p (type,TREE_TYPE (op1)))
-      || (TREE_CODE (op2) == SSA_NAME
-          && !types_compatible_p (type, TREE_TYPE (op2)))
-      || (op3 && TREE_CODE (op3) == SSA_NAME
-          && !types_compatible_p (type, TREE_TYPE (op3)))
-      || (op4 && TREE_CODE (op4) == SSA_NAME
-          && !types_compatible_p (type, TREE_TYPE (op4))))
-    {
-      if (dump_enabled_p ())
-        {
-          dump_printf_loc (MSG_NOTE, vect_location,
-                          "reduction: multiple types: operation type: "
-                          "%T, operands types: %T,%T",
-                          type,  TREE_TYPE (op1), TREE_TYPE (op2));
-          if (op3)
-           dump_printf (MSG_NOTE, ",%T", TREE_TYPE (op3));
-
-          if (op4)
-           dump_printf (MSG_NOTE, ",%T", TREE_TYPE (op4));
-          dump_printf (MSG_NOTE, "\n");
-        }
-
-      return NULL;
-    }
-
   /* Check whether it's ok to change the order of the computation.
      Generally, when vectorizing a reduction we change the order of the
      computation.  This may change the behavior of the program in some
@@ -2985,6 +2924,7 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
      vectorizing an outer-loop: the inner-loop is executed sequentially,
      and therefore vectorizing reductions in the inner-loop during
      outer-loop vectorization is safe.  */
+  tree type = TREE_TYPE (gimple_assign_lhs (def_stmt));
   if (STMT_VINFO_REDUC_TYPE (phi_info) == TREE_CODE_REDUCTION
       && needs_fold_left_reduction_p (type, code))
     STMT_VINFO_REDUC_TYPE (phi_info) = FOLD_LEFT_REDUCTION;
@@ -2993,38 +2933,19 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
      1) integer arithmetic and no trapv
      2) floating point arithmetic, and special flags permit this optimization
      3) nested cycle (i.e., outer loop vectorization).  */
+
+  /* Check for the simple case that one def is the reduction def,
+     defined by the PHI node.  */
   stmt_vec_info def1_info = loop_info->lookup_def (op1);
   stmt_vec_info def2_info = loop_info->lookup_def (op2);
-  if (code != COND_EXPR && !def1_info && !def2_info)
-    {
-      if (dump_enabled_p ())
-       report_vect_op (MSG_NOTE, def_stmt, "reduction: no defs for operands: ");
-      return NULL;
-    }
-
-  /* Check that one def is the reduction def, defined by PHI,
-     the other def is either defined in the loop ("vect_internal_def"),
-     or it's an induction (defined by a loop-header phi-node).  */
-
-  if (def2_info
-      && def2_info->stmt == phi
-      && (code == COND_EXPR
-         || !def1_info
-         || !flow_bb_inside_loop_p (loop, gimple_bb (def1_info->stmt))
-         || vect_valid_reduction_input_p (def1_info)))
+  if (def2_info && def2_info->stmt == phi)
     {
       STMT_VINFO_REDUC_IDX (def_stmt_info) = 1 + (code == COND_EXPR ? 1 : 0);
       if (dump_enabled_p ())
        report_vect_op (MSG_NOTE, def_stmt, "detected reduction: ");
       return def_stmt_info;
     }
-
-  if (def1_info
-      && def1_info->stmt == phi
-      && (code == COND_EXPR
-         || !def2_info
-         || !flow_bb_inside_loop_p (loop, gimple_bb (def2_info->stmt))
-         || vect_valid_reduction_input_p (def2_info)))
+  else if (def1_info && def1_info->stmt == phi)
     {
       STMT_VINFO_REDUC_IDX (def_stmt_info) = 0 + (code == COND_EXPR ? 1 : 0);
       if (dump_enabled_p ())
@@ -3032,7 +2953,8 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
       return def_stmt_info;
     }
 
-  /* Look for the expression computing latch_def from loop PHI result.  */
+  /* Look for the expression computing latch_def from then loop PHI result
+     in a way involving more than one stmt.  */
   auto_vec<std::pair<ssa_op_iter, use_operand_p> > path;
   if (check_reduction_path (vect_location, loop, phi, latch_def, code,
                            path))
@@ -5716,7 +5638,6 @@ vectorizable_reduction (stmt_vec_info stmt_info, slp_tree slp_node,
   enum vect_def_type dt, cond_reduc_dt = vect_unknown_def_type;
   stmt_vec_info cond_stmt_vinfo = NULL;
   tree scalar_type;
-  bool is_simple_use;
   int i;
   int ncopies;
   bool single_defuse_cycle = false;
@@ -5887,10 +5808,15 @@ vectorizable_reduction (stmt_vec_info stmt_info, slp_tree slp_node,
         continue;
 
       stmt_vec_info def_stmt_info;
-      is_simple_use = vect_is_simple_use (ops[i], loop_vinfo, &dts[i], &tem,
-                                         &def_stmt_info);
+      if (!vect_is_simple_use (ops[i], loop_vinfo, &dts[i], &tem,
+                              &def_stmt_info))
+       {
+         if (dump_enabled_p ())
+           dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+                            "use not simple.\n");
+         return false;
+       }
       dt = dts[i];
-      gcc_assert (is_simple_use);
       if (dt == vect_reduction_def
          && ops[i] == reduc_def)
        {
@@ -6000,6 +5926,15 @@ vectorizable_reduction (stmt_vec_info stmt_info, slp_tree slp_node,
          return false;
        }
 
+      /* When the condition uses the reduction value in the condition, fail.  */
+      if (reduc_index == 0)
+       {
+         if (dump_enabled_p ())
+           dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+                            "condition depends on previous iteration\n");
+         return false;
+       }
+
       if (direct_internal_fn_supported_p (IFN_FOLD_EXTRACT_LAST,
                                          vectype_in, OPTIMIZE_FOR_SPEED))
        {
index ff9863f290f281a88f6a74a90f267815430a7e05..e7255fb76bc12b9e43046d2891ddbd511412dbae 100644 (file)
@@ -10901,6 +10901,7 @@ vect_transform_stmt (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi,
   stmt_vec_info orig_stmt_info = vect_orig_stmt (stmt_info);
   stmt_vec_info reduc_info;
   if (STMT_VINFO_REDUC_DEF (orig_stmt_info)
+      && vect_stmt_to_vectorize (orig_stmt_info) == stmt_info
       && (reduc_info = info_for_reduction (orig_stmt_info))
       && STMT_VINFO_REDUC_TYPE (reduc_info) != FOLD_LEFT_REDUCTION
       && STMT_VINFO_REDUC_TYPE (reduc_info) != EXTRACT_LAST_REDUCTION)