[PATCH] Fix VTA ICE after fixup_noreturn_call (PR debug/48768)

[PATCH] Fix VTA ICE after fixup_noreturn_call (PR debug/48768)

[Jakub Jelinek]

Hi!

fixup_noreturn_call when removing LHS of a noreturn call does:
      /* We need to remove SSA name to avoid checking errors.
         All uses are dominated by the noreturn and thus will
         be removed afterwards.
         We proactively remove affected non-PHI statements to avoid
         fixup_cfg from trying to update them and crashing.  */
      if (TREE_CODE (op) == SSA_NAME)
        {
          use_operand_p use_p;
          imm_use_iterator iter;
          gimple use_stmt;
          bitmap_iterator bi;
          unsigned int bb_index;

          bitmap blocks = BITMAP_ALLOC (NULL);

          FOR_EACH_IMM_USE_STMT (use_stmt, iter, op)
            {
              if (gimple_code (use_stmt) != GIMPLE_PHI)
                bitmap_set_bit (blocks, gimple_bb (use_stmt)->index);
              else
                FOR_EACH_IMM_USE_ON_STMT (use_p, iter)
                  SET_USE (use_p, error_mark_node);
            }
which is fine because all the uses will be cleaned up RSN afterwards.
Unfortunately, during the cleanup VTA triggers and tries to add DEBUG exprs,
if there is a degenerate PHI which contains just error_mark_node set above
unfortunately that results in DEBUG something => <<< error >>>, which isn't
valid.

Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux,
ok for trunk/4.6?

2011-04-26  Jakub Jelinek  <jakub@redhat.com>

	PR debug/48768
	* tree-ssa.c (insert_debug_temp_for_var_def): If degenerate_phi_result
	is error_mark_node, set value to NULL.

	* gcc.dg/pr48768.c: New test.

--- gcc/tree-ssa.c.jj	2011-04-13 10:54:49.000000000 +0200
+++ gcc/tree-ssa.c	2011-04-26 10:56:28.000000000 +0200
@@ -352,6 +352,10 @@ insert_debug_temp_for_var_def (gimple_st
       value = degenerate_phi_result (def_stmt);
       if (value && walk_tree (&value, find_released_ssa_name, NULL, NULL))
 	value = NULL;
+      /* error_mark_node is what fixup_noreturn_call changes PHI arguments
+	 to.  */
+      else if (value == error_mark_node)
+	value = NULL;
     }
   else if (is_gimple_assign (def_stmt))
     {
--- gcc/testsuite/gcc.dg/pr48768.c.jj	2011-04-26 10:58:17.000000000 +0200
+++ gcc/testsuite/gcc.dg/pr48768.c	2011-04-26 10:57:30.000000000 +0200
@@ -0,0 +1,38 @@
+/* PR debug/48768 */
+/* { dg-do compile } */
+/* { dg-options "-O -fcompare-debug" } */
+
+int a, b;
+
+int
+bar (void)
+{
+  int i, j = 1;
+  for (i = 0; i != 10; i++)
+    {
+    lab:
+      if (i)
+	{
+	  int *k = &j;
+	}
+      else if (j)
+	goto lab;
+    }
+  return 1;
+}
+
+inline int
+foo (int x)
+{
+  unsigned int c = x;
+  int d = x;
+  if (bar ())
+    for (; c; c++)
+      while (x >= 0)
+	if (foo (d) >= 0)
+	  {
+	    d = bar ();
+	    a = b ? b : 1;
+	  }
+  return 0;
+}

	Jakub

Re: [PATCH] Fix VTA ICE after fixup_noreturn_call (PR debug/48768)

[Richard Guenther]

On Tue, Apr 26, 2011 at 3:08 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> fixup_noreturn_call when removing LHS of a noreturn call does:
>      /* We need to remove SSA name to avoid checking errors.
>         All uses are dominated by the noreturn and thus will
>         be removed afterwards.
>         We proactively remove affected non-PHI statements to avoid
>         fixup_cfg from trying to update them and crashing.  */
>      if (TREE_CODE (op) == SSA_NAME)
>        {
>          use_operand_p use_p;
>          imm_use_iterator iter;
>          gimple use_stmt;
>          bitmap_iterator bi;
>          unsigned int bb_index;
>
>          bitmap blocks = BITMAP_ALLOC (NULL);
>
>          FOR_EACH_IMM_USE_STMT (use_stmt, iter, op)
>            {
>              if (gimple_code (use_stmt) != GIMPLE_PHI)
>                bitmap_set_bit (blocks, gimple_bb (use_stmt)->index);
>              else
>                FOR_EACH_IMM_USE_ON_STMT (use_p, iter)
>                  SET_USE (use_p, error_mark_node);
>            }
> which is fine because all the uses will be cleaned up RSN afterwards.
> Unfortunately, during the cleanup VTA triggers and tries to add DEBUG exprs,
> if there is a degenerate PHI which contains just error_mark_node set above
> unfortunately that results in DEBUG something => <<< error >>>, which isn't
> valid.
>
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux,
> ok for trunk/4.6?

Ok.

Thanks,
Richard.

> 2011-04-26  Jakub Jelinek  <jakub@redhat.com>
>
>        PR debug/48768
>        * tree-ssa.c (insert_debug_temp_for_var_def): If degenerate_phi_result
>        is error_mark_node, set value to NULL.
>
>        * gcc.dg/pr48768.c: New test.
>
> --- gcc/tree-ssa.c.jj   2011-04-13 10:54:49.000000000 +0200
> +++ gcc/tree-ssa.c      2011-04-26 10:56:28.000000000 +0200
> @@ -352,6 +352,10 @@ insert_debug_temp_for_var_def (gimple_st
>       value = degenerate_phi_result (def_stmt);
>       if (value && walk_tree (&value, find_released_ssa_name, NULL, NULL))
>        value = NULL;
> +      /* error_mark_node is what fixup_noreturn_call changes PHI arguments
> +        to.  */
> +      else if (value == error_mark_node)
> +       value = NULL;
>     }
>   else if (is_gimple_assign (def_stmt))
>     {
> --- gcc/testsuite/gcc.dg/pr48768.c.jj   2011-04-26 10:58:17.000000000 +0200
> +++ gcc/testsuite/gcc.dg/pr48768.c      2011-04-26 10:57:30.000000000 +0200
> @@ -0,0 +1,38 @@
> +/* PR debug/48768 */
> +/* { dg-do compile } */
> +/* { dg-options "-O -fcompare-debug" } */
> +
> +int a, b;
> +
> +int
> +bar (void)
> +{
> +  int i, j = 1;
> +  for (i = 0; i != 10; i++)
> +    {
> +    lab:
> +      if (i)
> +       {
> +         int *k = &j;
> +       }
> +      else if (j)
> +       goto lab;
> +    }
> +  return 1;
> +}
> +
> +inline int
> +foo (int x)
> +{
> +  unsigned int c = x;
> +  int d = x;
> +  if (bar ())
> +    for (; c; c++)
> +      while (x >= 0)
> +       if (foo (d) >= 0)
> +         {
> +           d = bar ();
> +           a = b ? b : 1;
> +         }
> +  return 0;
> +}
>
>        Jakub
>