This is the mail archive of the binutils@sources.redhat.com mailing list for the binutils project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: PATCH: Change ia64 br relaxation failure message


On Thu, Feb 10, 2005 at 04:05:11PM -0800, H. J. Lu wrote:
> R_IA64_PCREL21B, R_IA64_PCREL21BI, R_IA64_PCREL21M and R_IA64_PCREL21F
> can fail if the input section is too big. Users get
> 
> foo.o(.text+0x1000072): In function `main':
> foo.c: relocation truncated to fit: PCREL21B against symbol `foo'
> defined in .text section in foo.o
> /usr/local/bin/ld: final link failed: Nonrepresentable section on
> output
> 
> and they may think it is a linker error. This patch changes it to
> 
> ./ld: foo.o: Can't relax br (PCREL21B) in section `.text' with size
> 0x10000b0 (> 0x1000000).
> ./ld: final link failed: Nonrepresentable section on output
> collect2: ld returned 1 exit status
> 

Hi Jim,

I just noticed a relaxation bug. We won't relax


foo:
	huge text
	br foo

since branch and target are in the same section. We can relax backward
branch in this case. Also we do &h->root without checking if h is NULL.
The updated patch fixes both problems.


H.J.
----
2005-02-10  H.J. Lu  <hongjiu.lu@intel.com>

	* elfxx-ia64.c (elfNN_ia64_relax_section): Allow relax
	backward branch in the same section.
	(elfNN_ia64_relocate_section): Inform users that the input
	section is too big to relax br instruction when overflow
	happens to R_IA64_PCREL21B, R_IA64_PCREL21BI, R_IA64_PCREL21M
	and R_IA64_PCREL21F. Check h before accessing &h->root.

--- bfd/elfxx-ia64.c.big	2005-02-11 11:16:33.000000000 -0800
+++ bfd/elfxx-ia64.c	2005-02-11 12:12:58.353313904 -0800
@@ -994,9 +994,10 @@ elfNN_ia64_relax_section (abfd, sec, lin
 	    }
 
 	  /* If the branch and target are in the same section, you've
-	     got one honking big section and we can't help you.  You'll
-	     get an error message later.  */
-	  if (tsec == sec)
+	     got one honking big section and we can't help you unless
+	     you are branching backwards.  You'll get an error message
+	     later.  */
+	  if (tsec == sec && toff > roff)
 	    continue;
 
 	  /* Look for an existing fixup to this address.  */
@@ -4542,13 +4543,37 @@ elfNN_ia64_relocate_section (output_bfd,
 		if (*name == '\0')
 		  name = bfd_section_name (input_bfd, sym_sec);
 	      }
-	    if (!(*info->callbacks->reloc_overflow) (info, &h->root,
-						     name, howto->name,
-						     (bfd_vma) 0,
-						     input_bfd,
-						     input_section,
-						     rel->r_offset))
-	      return FALSE;
+
+	    switch (r_type)
+	      {
+	      case R_IA64_PCREL21B:
+	      case R_IA64_PCREL21BI:
+	      case R_IA64_PCREL21M:
+	      case R_IA64_PCREL21F:
+		if (is_elf_hash_table (info->hash))
+		  {
+		    /* Relaxtion is always performed for ELF output.
+		       Overflow failures for those relocations mean
+		       that the section is too big to relax.  */
+		    (*_bfd_error_handler)
+		      (_("%B: Can't relax br (%s) to `%s' at 0x%lx in section `%A' with size 0x%lx (> 0x1000000)."),
+		       input_bfd, input_section, howto->name, name,
+		       rel->r_offset, input_section->size);
+		    break;
+		  }
+	      default:
+		if (!(*info->callbacks->reloc_overflow) (info,
+							 h ? &h->root : NULL,
+							 name,
+							 howto->name,
+							 (bfd_vma) 0,
+							 input_bfd,
+							 input_section,
+							 rel->r_offset))
+		  return FALSE;
+		break;
+	      }
+
 	    ret_val = FALSE;
 	  }
 	  break;


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]