IA64 linker bug: Linker failed to relax long call

IA64 linker bug: Linker failed to relax long call

[H. J. Lu]

See

http://sources.redhat.com/bugzilla/show_bug.cgi?id=240


H.J.

PATCH: IA64: Don't relax branch in .init/.fini sections.

[H. J. Lu]

On Sat, Jun 26, 2004 at 05:48:11PM -0700, H. J. Lu wrote:
> See
> 
> http://sources.redhat.com/bugzilla/show_bug.cgi?id=240
> 

Here is a patch.


H.J.
----
2004-06-26  H.J. Lu  <hongjiu.lu@intel.com>

	* elfxx-ia64.c (elfNN_ia64_relax_section): Don't relax branch
	in .init/.fini sections.

--- bfd/elfxx-ia64.c.brl	2004-06-24 08:49:54.000000000 -0700
+++ bfd/elfxx-ia64.c	2004-06-26 21:50:27.637103414 -0700
@@ -953,6 +953,18 @@ elfNN_ia64_relax_section (abfd, sec, lin
 	{
 	  bfd_signed_vma offset;
 
+	  /* We can't put a trampoline in a .init/.fini section. Issue
+	     an error.  */
+	  if (strcmp (sec->output_section->name, ".init") == 0
+	      || strcmp (sec->output_section->name, ".fini") == 0)
+	    {
+	      (*_bfd_error_handler)
+		(_("%s: Can't relax br in section `%s'. Please use brl."),
+		 bfd_archive_filename (sec->owner), sec->name);
+	      bfd_set_error (bfd_error_bad_value);
+	      goto error_return;
+	    }
+
 	  reladdr = (sec->output_section->vma
 		     + sec->output_offset
 		     + roff) & (bfd_vma) -4;

Re: PATCH: IA64: Don't relax branch in .init/.fini sections.

[Jim Wilson]

On Sat, 2004-06-26 at 22:53, H. J. Lu wrote:
> 	* elfxx-ia64.c (elfNN_ia64_relax_section): Don't relax branch
> 	in .init/.fini sections.

This looks OK to me.

You suggest using brl.  The gcc startup files deliberately use indirect
branches to avoid this problem.  We load the function address into a
register, and then branch on the register.  This also avoids use of brl
which is a slow on an Itanium1.  If we can ignore Itanium1, then we can
simplify the gcc startup code a little bit.

The error message doesn't give any hint about which instruction can't be
relocated.  Which means if you have lots of them you are screwed.  This
might be rare enough that we don't have to worry about it.  You didn't
give any info on how the problem was detected, so I can't tell if this
matters.
-- 
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com

Re: PATCH: IA64: Don't relax branch in .init/.fini sections.

[H. J. Lu]

On Tue, Jun 29, 2004 at 04:49:05PM -0700, Jim Wilson wrote:
> On Sat, 2004-06-26 at 22:53, H. J. Lu wrote:
> > 	* elfxx-ia64.c (elfNN_ia64_relax_section): Don't relax branch
> > 	in .init/.fini sections.
> 
> This looks OK to me.
> 
> You suggest using brl.  The gcc startup files deliberately use indirect

I will change that.

> branches to avoid this problem.  We load the function address into a
> register, and then branch on the register.  This also avoids use of brl
> which is a slow on an Itanium1.  If we can ignore Itanium1, then we can
> simplify the gcc startup code a little bit.

I think we can do it for gcc 3.4 and above.

> 
> The error message doesn't give any hint about which instruction can't be
> relocated.  Which means if you have lots of them you are screwed.  This

I will see what I can do.

> might be rare enough that we don't have to worry about it.  You didn't
> give any info on how the problem was detected, so I can't tell if this
> matters.


It came from icc. When I used icc to build a big application, linker
failed to issue an error. I got run-time failure instead.

Thanks.


H.J.

Re: PATCH: IA64: Don't relax branch in .init/.fini sections.

[H. J. Lu]

[-- Attachment #1: Type: text/plain, Size: 919 bytes --]

On Tue, Jun 29, 2004 at 04:49:05PM -0700, Jim Wilson wrote:
> On Sat, 2004-06-26 at 22:53, H. J. Lu wrote:
> > 	* elfxx-ia64.c (elfNN_ia64_relax_section): Don't relax branch
> > 	in .init/.fini sections.
> 
> This looks OK to me.
> 
> You suggest using brl.  The gcc startup files deliberately use indirect
> branches to avoid this problem.  We load the function address into a
> register, and then branch on the register.  This also avoids use of brl
> which is a slow on an Itanium1.  If we can ignore Itanium1, then we can
> simplify the gcc startup code a little bit.
> 
> The error message doesn't give any hint about which instruction can't be
> relocated.  Which means if you have lots of them you are screwed.  This
> might be rare enough that we don't have to worry about it.  You didn't
> give any info on how the problem was detected, so I can't tell if this
> matters.

I will check in this shortly.


H.J.

[-- Attachment #2: bfd-ia64-brl-2.patch --]
[-- Type: text/plain, Size: 948 bytes --]

2004-06-29  H.J. Lu  <hongjiu.lu@intel.com>

	PR 240
	* elfxx-ia64.c (elfNN_ia64_relax_section): Don't relax branch
	in .init/.fini sections.

--- bfd/elfxx-ia64.c.brl	2004-06-29 14:41:42.833932627 -0700
+++ bfd/elfxx-ia64.c	2004-06-29 18:10:44.212915584 -0700
@@ -953,6 +953,19 @@ elfNN_ia64_relax_section (abfd, sec, lin
 	{
 	  bfd_signed_vma offset;
 
+	  /* We can't put a trampoline in a .init/.fini section. Issue
+	     an error.  */
+	  if (strcmp (sec->output_section->name, ".init") == 0
+	      || strcmp (sec->output_section->name, ".fini") == 0)
+	    {
+	      (*_bfd_error_handler)
+		(_("%s: Can't relax br at 0x%lx in section `%s'. Please use brl or indirect branch."),
+		 bfd_archive_filename (sec->owner),
+		 (unsigned long) roff, sec->name);
+	      bfd_set_error (bfd_error_bad_value);
+	      goto error_return;
+	    }
+
 	  reladdr = (sec->output_section->vma
 		     + sec->output_offset
 		     + roff) & (bfd_vma) -4;