Why does ELF/alpha put none-text relocations in .rela.text?

Why does ELF/alpha put none-text relocations in .rela.text?

[H . J . Lu]

When you do

# gcc -c foo.c

on ELF/alpha, as will put relocations for ".debug_line" in
.rela.text. When you do

# strip --strip-debug foo.o
# readelf -r foo.o

you will get

 00000000  200569c000000006 R_ALPHA_GPDISP        bad symbol index: 200569c0

How should we fix it?


H.J.

Re: Why does ELF/alpha put none-text relocations in .rela.text?

[Richard Henderson]

On Mon, Aug 20, 2001 at 11:44:19AM -0700, H . J . Lu wrote:
>  00000000  200569c000000006 R_ALPHA_GPDISP        bad symbol index: 200569c0
> 
> How should we fix it?

The symbol for GPDISP is completely irrelevant.  Only the offset
is meaningful.  Use the current section symbol rather than a
random section symbol.


r~

Re: Why does ELF/alpha put none-text relocations in .rela.text?

[H . J . Lu]

On Mon, Aug 20, 2001 at 03:03:28PM -0700, Richard Henderson wrote:
> On Mon, Aug 20, 2001 at 11:44:19AM -0700, H . J . Lu wrote:
> >  00000000  200569c000000006 R_ALPHA_GPDISP        bad symbol index: 200569c0
> > 
> > How should we fix it?
> 
> The symbol for GPDISP is completely irrelevant.  Only the offset
> is meaningful.  Use the current section symbol rather than a
> random section symbol.

But it is a bad idea to use the debug section symbol since it may
be removed.


H.J.

Re: Why does ELF/alpha put none-text relocations in .rela.text?

[H . J . Lu]

On Mon, Aug 20, 2001 at 03:03:28PM -0700, Richard Henderson wrote:
> On Mon, Aug 20, 2001 at 11:44:19AM -0700, H . J . Lu wrote:
> >  00000000  200569c000000006 R_ALPHA_GPDISP        bad symbol index: 200569c0
> > 
> > How should we fix it?
> 
> The symbol for GPDISP is completely irrelevant.  Only the offset
> is meaningful.  Use the current section symbol rather than a
> random section symbol.
> 

But we are using a random section symbol since we never reset now_seg.
This patch seems to work for me. Alan, do you have any comments?

Thanks.


H.J.
----
2001-08-20  H.J. Lu  <hjl@gnu.org>

	* write.c (write_relocs): Reset the current section.

Index: write.c
===================================================================
RCS file: /work/cvs/gnu/binutils/gas/write.c,v
retrieving revision 1.26
diff -u -p -r1.26 write.c
--- write.c	2001/08/03 15:55:50	1.26
+++ write.c	2001/08/20 23:00:50
@@ -958,6 +958,8 @@ write_relocs (abfd, sec, xxx)
   if (seginfo == NULL)
     return;
 
+  subseg_change (sec, 0);
+
   fixup_segment (seginfo->fix_root, sec);
 
   n = 0;

PATCH: Avoid setting the bogus section symbol index.

[H . J . Lu]

On Mon, Aug 20, 2001 at 04:05:15PM -0700, H . J . Lu wrote:
> On Mon, Aug 20, 2001 at 03:03:28PM -0700, Richard Henderson wrote:
> > On Mon, Aug 20, 2001 at 11:44:19AM -0700, H . J . Lu wrote:
> > >  00000000  200569c000000006 R_ALPHA_GPDISP        bad symbol index: 200569c0
> > > 
> > > How should we fix it?
> > 
> > The symbol for GPDISP is completely irrelevant.  Only the offset
> > is meaningful.  Use the current section symbol rather than a
> > random section symbol.
> > 
> 

Here is a patch to avoid setting the bogus section symbol index. Now,
I am getting:

/export/build/gnu/binutils/build-alpha-linux/binutils/strip-new --strip-debug foo.o -o strip.o
BFD: strip.o: symbol `.debug_line' required but not present

I will check it in shortly if there is no objection.


H.J.
----
2001-08-20  H.J. Lu  <hjl@gnu.org>

	* elf-bfd.h (elf_obj_tdata): Add num_section_syms.
	(elf_num_section_syms): New for num_section_syms.

	* elf.c (elf_map_symbols): Set num_section_syms.
	(_bfd_elf_symbol_from_bfd_symbol): Check num_section_syms for
	the section symbol index.

Index: elf-bfd.h
===================================================================
RCS file: /work/cvs/gnu/binutils/bfd/elf-bfd.h,v
retrieving revision 1.15
diff -u -p -r1.15 elf-bfd.h
--- elf-bfd.h	2001/08/17 17:56:19	1.15
+++ elf-bfd.h	2001/08/20 23:42:08
@@ -821,6 +821,7 @@ struct elf_obj_tdata
   struct bfd_strtab_hash *strtab_ptr;
   int num_locals;
   int num_globals;
+  int num_section_syms;
   asymbol **section_syms;		/* STT_SECTION symbols for each section */
   Elf_Internal_Shdr symtab_hdr;
   Elf_Internal_Shdr shstrtab_hdr;
@@ -962,6 +963,7 @@ struct elf_obj_tdata
 #define elf_num_locals(bfd)	(elf_tdata(bfd) -> num_locals)
 #define elf_num_globals(bfd)	(elf_tdata(bfd) -> num_globals)
 #define elf_section_syms(bfd)	(elf_tdata(bfd) -> section_syms)
+#define elf_num_section_syms(bfd) (elf_tdata(bfd) -> num_section_syms)
 #define core_prpsinfo(bfd)	(elf_tdata(bfd) -> prpsinfo)
 #define core_prstatus(bfd)	(elf_tdata(bfd) -> prstatus)
 #define elf_gp(bfd)		(elf_tdata(bfd) -> gp)
Index: elf.c
===================================================================
RCS file: /work/cvs/gnu/binutils/bfd/elf.c,v
retrieving revision 1.60
diff -u -p -r1.60 elf.c
--- elf.c	2001/08/17 17:56:19	1.60
+++ elf.c	2001/08/20 23:42:08
@@ -2092,6 +2092,7 @@ elf_map_symbols (abfd)
   if (sect_syms == NULL)
     return false;
   elf_section_syms (abfd) = sect_syms;
+  elf_num_section_syms (abfd) = max_index;
 
   for (idx = 0; idx < symcount; idx++)
     {
@@ -3652,7 +3653,8 @@ _bfd_elf_symbol_from_bfd_symbol (abfd, a
 	indx = asym_ptr->section->output_section->index;
       else
 	indx = asym_ptr->section->index;
-      if (elf_section_syms (abfd)[indx])
+      if (indx < elf_num_section_syms (abfd)
+	  && elf_section_syms (abfd)[indx])
 	asym_ptr->udata.i = elf_section_syms (abfd)[indx]->udata.i;
     }
 

Re: Why does ELF/alpha put none-text relocations in .rela.text?

[Richard Henderson]

On Mon, Aug 20, 2001 at 04:05:15PM -0700, H . J . Lu wrote:
> But we are using a random section symbol since we never reset now_seg.

Um, no, that's just as broken.



r~


	* config/tc-alpha.h (MD_APPLY_FIX3): Define.
	* config/tc-alpha.c (md_apply_fix3): Rename from md_apply_fix;
	use seg parameter instead of now_seg.

Index: tc-alpha.c
===================================================================
RCS file: /cvs/src/src/gas/config/tc-alpha.c,v
retrieving revision 1.24
diff -c -p -d -u -r1.24 tc-alpha.c
--- tc-alpha.c	2001/06/24 16:43:31	1.24
+++ tc-alpha.c	2001/08/21 00:08:15
@@ -1159,9 +1159,10 @@ md_pcrel_from (fixP)
    GPDISP.  */
 
 int
-md_apply_fix (fixP, valueP)
+md_apply_fix3 (fixP, valueP, seg)
      fixS *fixP;
      valueT *valueP;
+     segT seg;
 {
   char * const fixpos = fixP->fx_frag->fr_literal + fixP->fx_where;
   valueT value = *valueP;
@@ -1196,7 +1197,7 @@ md_apply_fix (fixP, valueP)
 #endif
 
     do_reloc_gp:
-      fixP->fx_addsy = section_symbol (now_seg);
+      fixP->fx_addsy = section_symbol (seg);
       md_number_to_chars (fixpos, value, 2);
       break;
 
Index: tc-alpha.h
===================================================================
RCS file: /cvs/src/src/gas/config/tc-alpha.h,v
retrieving revision 1.7
diff -c -p -d -u -r1.7 tc-alpha.h
--- tc-alpha.h	2001/07/23 14:02:12	1.7
+++ tc-alpha.h	2001/08/21 00:08:15
@@ -49,6 +49,7 @@ extern valueT alpha_gp_value;
 #define TC_FORCE_RELOCATION(FIXP)	alpha_force_relocation (FIXP)
 #define tc_fix_adjustable(FIXP)		alpha_fix_adjustable (FIXP)
 #define RELOC_REQUIRES_SYMBOL
+#define MD_APPLY_FIX3
 
 /* This expression evaluates to false if the relocation is for a local
    object for which we still want to do the relocation at runtime.

Re: Why does ELF/alpha put none-text relocations in .rela.text?

[H . J . Lu]

On Mon, Aug 20, 2001 at 11:44:19AM -0700, H . J . Lu wrote:
> When you do
> 
> # gcc -c foo.c
> 
> on ELF/alpha, as will put relocations for ".debug_line" in
> .rela.text. When you do
> 
> # strip --strip-debug foo.o
> # readelf -r foo.o
> 
> you will get
> 
>  00000000  200569c000000006 R_ALPHA_GPDISP        bad symbol index: 200569c0
> 
> How should we fix it?

I think the real problem is ELF/alpha abuses ELF64_R_SYM for the GPDISP
relocations where the symbol index is not really an index into the
symbol table. But the generic part of the ELF linker in BFD assumes
ELF64_R_SYM is a real symbol index. I will see what I can do.


H.J.