Patchwork Fix PR middle-end/49732

login
register
mail settings
Submitter Eric Botcazou
Date July 14, 2011, 9:44 p.m.
Message ID <201107142344.33806.ebotcazou@adacore.com>
Download mbox | patch
Permalink /patch/104730/
State New
Headers show

Comments

Eric Botcazou - July 14, 2011, 9:44 p.m.
Hi,

this is a regression present on mainline and 4.6 branch.  The compiler crashes 
during gimplification because there is a COMPOUND_EXPR shared between the 
TYPE_SIZE and TYPE_SIZE_UNIT expressions of an array type.  Now this isn't 
supposed to happen because we run an unsharing pass before gimplification.

The problem here is that we have a forward declaration (DECL_EXPR) of a pointer 
type to the array type.  So, during the marking phase of the unsharing pass, 
the array type gets marked as visited through the pointer, which prevents it 
from being walked during the same phase when its own DECL_EXPR is processed.

This pointer/pointed-to type business is an old pattern.  Five years ago, 
Olivier changed gimplify_type_sizes like so:

2006-10-06  Olivier Hainque  <hainque@adacore.com>

	* gimplify.c (gimplify_type_sizes) [POINTER_TYPE, REFERENCE_TYPE]:
	Don't recurse on the pointed-to type.

because of a related problem.  It turns out that we need the same change in the 
DECL_EXPR case of walk_tree_1 to fix the bug at hand, which is sort of logical 
as there is a strong relationship between them:

    case DECL_EXPR:
      /* If this is a TYPE_DECL, walk into the fields of the type that it's
	 defining.  We only want to walk into these fields of a type in this
	 case and not in the general case of a mere reference to the type.

	 The criterion is as follows: if the field can be an expression, it
	 must be walked only here.  This should be in keeping with the fields
	 that are directly gimplified in gimplify_type_sizes in order for the
	 mark/copy-if-shared/unmark machinery of the gimplifier to work with
	 variable-sized types.

	 Note that DECLs get walked as part of processing the BIND_EXPR.  */


Tested on x86_64-suse-linux, OK for mainline and 4.6 branch?


2011-07-14  Eric Botcazou  <ebotcazou@adacore.com>

	PR middle-end/49732
	* tree.c (walk_tree_1) <DECL_EXPR>: Do not walk a pointed-to type.


2011-07-14  Eric Botcazou  <ebotcazou@adacore.com>

	* gnat.dg/pointer_controlled.adb: New test.
Richard Guenther - July 16, 2011, 11:20 a.m.
On Thu, Jul 14, 2011 at 11:44 PM, Eric Botcazou <ebotcazou@adacore.com> wrote:
> Hi,
>
> this is a regression present on mainline and 4.6 branch.  The compiler crashes
> during gimplification because there is a COMPOUND_EXPR shared between the
> TYPE_SIZE and TYPE_SIZE_UNIT expressions of an array type.  Now this isn't
> supposed to happen because we run an unsharing pass before gimplification.
>
> The problem here is that we have a forward declaration (DECL_EXPR) of a pointer
> type to the array type.  So, during the marking phase of the unsharing pass,
> the array type gets marked as visited through the pointer, which prevents it
> from being walked during the same phase when its own DECL_EXPR is processed.
>
> This pointer/pointed-to type business is an old pattern.  Five years ago,
> Olivier changed gimplify_type_sizes like so:
>
> 2006-10-06  Olivier Hainque  <hainque@adacore.com>
>
>        * gimplify.c (gimplify_type_sizes) [POINTER_TYPE, REFERENCE_TYPE]:
>        Don't recurse on the pointed-to type.
>
> because of a related problem.  It turns out that we need the same change in the
> DECL_EXPR case of walk_tree_1 to fix the bug at hand, which is sort of logical
> as there is a strong relationship between them:
>
>    case DECL_EXPR:
>      /* If this is a TYPE_DECL, walk into the fields of the type that it's
>         defining.  We only want to walk into these fields of a type in this
>         case and not in the general case of a mere reference to the type.
>
>         The criterion is as follows: if the field can be an expression, it
>         must be walked only here.  This should be in keeping with the fields
>         that are directly gimplified in gimplify_type_sizes in order for the
>         mark/copy-if-shared/unmark machinery of the gimplifier to work with
>         variable-sized types.
>
>         Note that DECLs get walked as part of processing the BIND_EXPR.  */
>
>
> Tested on x86_64-suse-linux, OK for mainline and 4.6 branch?

Ok.

Thanks,
Richard.

>
> 2011-07-14  Eric Botcazou  <ebotcazou@adacore.com>
>
>        PR middle-end/49732
>        * tree.c (walk_tree_1) <DECL_EXPR>: Do not walk a pointed-to type.
>
>
> 2011-07-14  Eric Botcazou  <ebotcazou@adacore.com>
>
>        * gnat.dg/pointer_controlled.adb: New test.
>
>
> --
> Eric Botcazou
>

Patch

Index: tree.c
===================================================================
--- tree.c	(revision 176261)
+++ tree.c	(working copy)
@@ -10596,9 +10596,14 @@  walk_tree_1 (tree *tp, walk_tree_fn func
 	  if (result || !walk_subtrees)
 	    return result;
 
-	  result = walk_type_fields (*type_p, func, data, pset, lh);
-	  if (result)
-	    return result;
+	  /* But do not walk a pointed-to type since it may itself need to
+	     be walked in the declaration case if it isn't anonymous.  */
+	  if (!POINTER_TYPE_P (*type_p))
+	    {
+	      result = walk_type_fields (*type_p, func, data, pset, lh);
+	      if (result)
+		return result;
+	    }
 
 	  /* If this is a record type, also walk the fields.  */
 	  if (RECORD_OR_UNION_TYPE_P (*type_p))