[PATCH] Fix reassoc update_range_test (PR tree-optimization/62209)

[PATCH] Fix reassoc update_range_test (PR tree-optimization/62209)

[Jakub Jelinek]

Hi!

With LTO and -O0 initial compilation, -O1 linking we at some point used to
add stmts normal after a PHI, which is invalid.
This patch fixes that by inserting after labels instead.
No testcase provided, as the issue is latent anyway and I don't know how to
write lto testcases that use different options initially from linking
options.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2015-02-13  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/62209
	* tree-ssa-reassoc.c (update_range_test): If stmt is a PHI and
	op == range->exp, insert seq and gimplified code after labels
	instead of after the phi.

--- gcc/tree-ssa-reassoc.c.jj	2015-02-13 15:02:03.000000000 +0100
+++ gcc/tree-ssa-reassoc.c	2015-02-13 16:52:05.113232491 +0100
@@ -2160,10 +2160,18 @@ update_range_test (struct range_entry *r
 
   tem = fold_convert_loc (loc, optype, tem);
   gsi = gsi_for_stmt (stmt);
+  unsigned int uid = gimple_uid (stmt);
   /* In rare cases range->exp can be equal to lhs of stmt.
      In that case we have to insert after the stmt rather then before
-     it.  */
-  if (op == range->exp)
+     it.  If stmt is a PHI, insert it at the start of the basic block.  */
+  if (op != range->exp)
+    {
+      gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT);
+      tem = force_gimple_operand_gsi (&gsi, tem, true, NULL_TREE, true,
+				      GSI_SAME_STMT);
+      gsi_prev (&gsi);
+    }
+  else if (gimple_code (stmt) != GIMPLE_PHI)
     {
       gsi_insert_seq_after (&gsi, seq, GSI_CONTINUE_LINKING);
       tem = force_gimple_operand_gsi (&gsi, tem, true, NULL_TREE, false,
@@ -2171,16 +2179,32 @@ update_range_test (struct range_entry *r
     }
   else
     {
+      gsi = gsi_after_labels (gimple_bb (stmt));
+      if (!gsi_end_p (gsi))
+	uid = gimple_uid (gsi_stmt (gsi));
+      else
+	{
+	  gsi = gsi_start_bb (gimple_bb (stmt));
+	  uid = 1;
+	  while (!gsi_end_p (gsi))
+	    {
+	      uid = gimple_uid (gsi_stmt (gsi));
+	      gsi_next (&gsi);
+	    }
+	}
       gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT);
       tem = force_gimple_operand_gsi (&gsi, tem, true, NULL_TREE, true,
 				      GSI_SAME_STMT);
-      gsi_prev (&gsi);
+      if (gsi_end_p (gsi))
+	gsi = gsi_last_bb (gimple_bb (stmt));
+      else
+	gsi_prev (&gsi);
     }
   for (; !gsi_end_p (gsi); gsi_prev (&gsi))
     if (gimple_uid (gsi_stmt (gsi)))
       break;
     else
-      gimple_set_uid (gsi_stmt (gsi), gimple_uid (stmt));
+      gimple_set_uid (gsi_stmt (gsi), uid);
 
   oe->op = tem;
   range->exp = exp;

	Jakub

Re: [PATCH] Fix reassoc update_range_test (PR tree-optimization/62209)

[Richard Biener]

On February 13, 2015 8:48:25 PM CET, Jakub Jelinek <jakub@redhat.com> wrote:
>Hi!
>
>With LTO and -O0 initial compilation, -O1 linking we at some point used
>to
>add stmts normal after a PHI, which is invalid.
>This patch fixes that by inserting after labels instead.
>No testcase provided, as the issue is latent anyway and I don't know
>how to
>write lto testcases that use different options initially from linking
>options.
>
>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

OK.

Thanks,
Richard.

>2015-02-13  Jakub Jelinek  <jakub@redhat.com>
>
>	PR tree-optimization/62209
>	* tree-ssa-reassoc.c (update_range_test): If stmt is a PHI and
>	op == range->exp, insert seq and gimplified code after labels
>	instead of after the phi.
>
>--- gcc/tree-ssa-reassoc.c.jj	2015-02-13 15:02:03.000000000 +0100
>+++ gcc/tree-ssa-reassoc.c	2015-02-13 16:52:05.113232491 +0100
>@@ -2160,10 +2160,18 @@ update_range_test (struct range_entry *r
> 
>   tem = fold_convert_loc (loc, optype, tem);
>   gsi = gsi_for_stmt (stmt);
>+  unsigned int uid = gimple_uid (stmt);
>   /* In rare cases range->exp can be equal to lhs of stmt.
>      In that case we have to insert after the stmt rather then before
>-     it.  */
>-  if (op == range->exp)
>+     it.  If stmt is a PHI, insert it at the start of the basic block.
> */
>+  if (op != range->exp)
>+    {
>+      gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT);
>+      tem = force_gimple_operand_gsi (&gsi, tem, true, NULL_TREE,
>true,
>+				      GSI_SAME_STMT);
>+      gsi_prev (&gsi);
>+    }
>+  else if (gimple_code (stmt) != GIMPLE_PHI)
>     {
>       gsi_insert_seq_after (&gsi, seq, GSI_CONTINUE_LINKING);
>     tem = force_gimple_operand_gsi (&gsi, tem, true, NULL_TREE, false,
>@@ -2171,16 +2179,32 @@ update_range_test (struct range_entry *r
>     }
>   else
>     {
>+      gsi = gsi_after_labels (gimple_bb (stmt));
>+      if (!gsi_end_p (gsi))
>+	uid = gimple_uid (gsi_stmt (gsi));
>+      else
>+	{
>+	  gsi = gsi_start_bb (gimple_bb (stmt));
>+	  uid = 1;
>+	  while (!gsi_end_p (gsi))
>+	    {
>+	      uid = gimple_uid (gsi_stmt (gsi));
>+	      gsi_next (&gsi);
>+	    }
>+	}
>       gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT);
>      tem = force_gimple_operand_gsi (&gsi, tem, true, NULL_TREE, true,
> 				      GSI_SAME_STMT);
>-      gsi_prev (&gsi);
>+      if (gsi_end_p (gsi))
>+	gsi = gsi_last_bb (gimple_bb (stmt));
>+      else
>+	gsi_prev (&gsi);
>     }
>   for (; !gsi_end_p (gsi); gsi_prev (&gsi))
>     if (gimple_uid (gsi_stmt (gsi)))
>       break;
>     else
>-      gimple_set_uid (gsi_stmt (gsi), gimple_uid (stmt));
>+      gimple_set_uid (gsi_stmt (gsi), uid);
> 
>   oe->op = tem;
>   range->exp = exp;
>
>	Jakub