public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH] Fix gimplify_parameters (PR middle-end/36125)
@ 2008-11-11 13:20 Jakub Jelinek
  2008-11-11 15:37 ` Richard Guenther
  0 siblings, 1 reply; 2+ messages in thread
From: Jakub Jelinek @ 2008-11-11 13:20 UTC (permalink / raw)
  To: gcc-patches

Hi!

If gimplify_parameters for callee copied parameters copies a PARM_DECL
into a local VAR_DECL, which is used by all other references in the function
instead of the PARM_DECL and the PARM_DECL has TREE_ADDRESSABLE flag,
invalid GIMPLE is emitted.  The TREE_ADDRESSABLE flag in this case can
mean just that the address of the parameter has been taken somewhere within
the function (PARM_DECLs with TREE_ADDRESSABLE types aren't callee copied),
so very likely during gimplification the local copy VAR_DECL is marked
TREE_ADDRESSABLE as well, and assigning from TREE_ADDRESSABLE PARM_DECL
into TREE_ADDRESSABLE var directly is not valid in GIMPLE.

IMHO we don't need to mark the PARM_DECL as TREE_ADDRESSABLE, its
address is certainly not taken anywhere in the function (as they take
address of the local copy instead).  So the following patch just moves
the TREE_ADDRESSABLE flag from the PARM_DECL to the VAR_DECL that is
stored in DECL_VALUE_EXPR of the PARM_DECL.
I've bootstrapped/regtested this on x86_64-linux, John, can you please
bootstrap it on hppa?  Ok for trunk?

2008-11-11  Jakub Jelinek  <jakub@redhat.com>

	PR middle-end/36125
	* function.c (gimplify_parameters): For callee copies parameters,
	move TREE_ADDRESSABLE flag from the PARM_DECL to the local copy.

	* gcc.c-torture/compile/pr36125.c: New test.

--- gcc/function.c.jj	2008-10-31 22:28:21.000000000 +0100
+++ gcc/function.c	2008-11-11 11:09:53.000000000 +0100
@@ -3265,6 +3265,14 @@ gimplify_parameters (void)
 		{
 		  local = create_tmp_var (type, get_name (parm));
 		  DECL_IGNORED_P (local) = 0;
+		  /* If PARM was addressable, move that flag over
+		     to the local copy, as its address will be taken,
+		     not the PARMs.  */
+		  if (TREE_ADDRESSABLE (parm))
+		    {
+		      TREE_ADDRESSABLE (parm) = 0;
+		      TREE_ADDRESSABLE (local) = 1;
+		    }
 		}
 	      else
 		{
--- gcc/testsuite/gcc.c-torture/compile/pr36125.c.jj	2008-11-11 11:11:51.000000000 +0100
+++ gcc/testsuite/gcc.c-torture/compile/pr36125.c	2008-11-11 11:10:34.000000000 +0100
@@ -0,0 +1,10 @@
+/* PR middle-end/36125 */
+
+extern void bar (long double *);
+
+int
+foo (long double x)
+{
+  bar (&x);
+  return 0;
+}

	Jakub

^ permalink raw reply	[flat|nested] 2+ messages in thread

* Re: [PATCH] Fix gimplify_parameters (PR middle-end/36125)
  2008-11-11 13:20 [PATCH] Fix gimplify_parameters (PR middle-end/36125) Jakub Jelinek
@ 2008-11-11 15:37 ` Richard Guenther
  0 siblings, 0 replies; 2+ messages in thread
From: Richard Guenther @ 2008-11-11 15:37 UTC (permalink / raw)
  To: Jakub Jelinek; +Cc: gcc-patches

On Tue, Nov 11, 2008 at 7:06 AM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> If gimplify_parameters for callee copied parameters copies a PARM_DECL
> into a local VAR_DECL, which is used by all other references in the function
> instead of the PARM_DECL and the PARM_DECL has TREE_ADDRESSABLE flag,
> invalid GIMPLE is emitted.  The TREE_ADDRESSABLE flag in this case can
> mean just that the address of the parameter has been taken somewhere within
> the function (PARM_DECLs with TREE_ADDRESSABLE types aren't callee copied),
> so very likely during gimplification the local copy VAR_DECL is marked
> TREE_ADDRESSABLE as well, and assigning from TREE_ADDRESSABLE PARM_DECL
> into TREE_ADDRESSABLE var directly is not valid in GIMPLE.
>
> IMHO we don't need to mark the PARM_DECL as TREE_ADDRESSABLE, its
> address is certainly not taken anywhere in the function (as they take
> address of the local copy instead).  So the following patch just moves
> the TREE_ADDRESSABLE flag from the PARM_DECL to the VAR_DECL that is
> stored in DECL_VALUE_EXPR of the PARM_DECL.
> I've bootstrapped/regtested this on x86_64-linux, John, can you please
> bootstrap it on hppa?  Ok for trunk?

Makes sense.  Ok if it works on hppa.

Thanks,
Richard.

> 2008-11-11  Jakub Jelinek  <jakub@redhat.com>
>
>        PR middle-end/36125
>        * function.c (gimplify_parameters): For callee copies parameters,
>        move TREE_ADDRESSABLE flag from the PARM_DECL to the local copy.
>
>        * gcc.c-torture/compile/pr36125.c: New test.
>
> --- gcc/function.c.jj   2008-10-31 22:28:21.000000000 +0100
> +++ gcc/function.c      2008-11-11 11:09:53.000000000 +0100
> @@ -3265,6 +3265,14 @@ gimplify_parameters (void)
>                {
>                  local = create_tmp_var (type, get_name (parm));
>                  DECL_IGNORED_P (local) = 0;
> +                 /* If PARM was addressable, move that flag over
> +                    to the local copy, as its address will be taken,
> +                    not the PARMs.  */
> +                 if (TREE_ADDRESSABLE (parm))
> +                   {
> +                     TREE_ADDRESSABLE (parm) = 0;
> +                     TREE_ADDRESSABLE (local) = 1;
> +                   }
>                }
>              else
>                {
> --- gcc/testsuite/gcc.c-torture/compile/pr36125.c.jj    2008-11-11 11:11:51.000000000 +0100
> +++ gcc/testsuite/gcc.c-torture/compile/pr36125.c       2008-11-11 11:10:34.000000000 +0100
> @@ -0,0 +1,10 @@
> +/* PR middle-end/36125 */
> +
> +extern void bar (long double *);
> +
> +int
> +foo (long double x)
> +{
> +  bar (&x);
> +  return 0;
> +}
>
>        Jakub
>

^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2008-11-11 15:35 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2008-11-11 13:20 [PATCH] Fix gimplify_parameters (PR middle-end/36125) Jakub Jelinek
2008-11-11 15:37 ` Richard Guenther

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).