sh-linux patch for 32-bit relative relocation

sh-linux patch for 32-bit relative relocation

[kaz Kojima]

Hello,

Here is a patch for sh-linux to handle the substraction of two
symbols in different sections correctly. This is essentially same
with the old patch
  http://sources.redhat.com/ml/binutils/2001-01/msg00071.html
which had a possible ABI problem in non linux targets, but is now
organized as a linux specific patch which doesn't affect non linux
targets. Formally the change for gas/tc-sh.c affects all sh targets,
but the effect of this change is ignored by non linux targets since
src_mask of HOWTO (R_SH_REL32,... is set to 0 in such targets.
I think that almost all sh-linux developpers are using this patch
successfully and enough test is done. Without this, we can't get
working shared libraries for stdc++, X and many other packages.

Regards,
	kaz
--
2001-09-11  Kaz Kojima  <kkojima@rr.iij4u.or.jp>

	* bfd/elf32-sh-lin.c (LINUX_ABI): Defined.
	* bfd/elf32-sh.c (sh_elf_howto_table): Don't mask source
	R_SH_REL32 relocation in linux case.
	(sh_elf_relocate_section): Use in-place value in computing
	addend for R_SH_REL32 relocation in linux case.

Index: bfd/elf32-sh-lin.c
===================================================================
RCS file: /cvs/src/src/bfd/elf32-sh-lin.c,v
retrieving revision 1.1
diff -u -r1.1 elf32-sh-lin.c
--- elf32-sh-lin.c	2000/09/07 04:23:43	1.1
+++ elf32-sh-lin.c	2001/08/22 01:26:50
@@ -25,5 +25,6 @@
 #define ELF_MACHINE_CODE EM_SH
 #define ELF_MAXPAGESIZE 0x10000
 #define elf_symbol_leading_char 0
+#define LINUX_ABI 1
 
 #include "elf32-sh.c"
Index: bfd/elf32-sh.c
===================================================================
RCS file: /cvs/src/src/bfd/elf32-sh.c,v
retrieving revision 1.29
diff -u -r1.29 elf32-sh.c
--- elf32-sh.c	2001/05/23 16:38:36	1.29
+++ elf32-sh.c	2001/08/22 01:27:09
@@ -112,6 +112,7 @@
 	 false),		/* pcrel_offset */
 
   /* 32 bit PC relative relocation.  */
+#if defined(LINUX_ABI)
   HOWTO (R_SH_REL32,		/* type */
 	 0,			/* rightshift */
 	 2,			/* size (0 = byte, 1 = short, 2 = long) */
@@ -121,10 +122,25 @@
 	 complain_overflow_signed, /* complain_on_overflow */
 	 sh_elf_ignore_reloc,	/* special_function */
 	 "R_SH_REL32",		/* name */
+	 true,			/* partial_inplace */
+	 0xffffffff,		/* src_mask */
+	 0xffffffff,		/* dst_mask */
+	 true),			/* pcrel_offset */
+#else
+  HOWTO (R_SH_REL32,		/* type */
+	 0,			/* rightshift */
+	 2,			/* size (0 = byte, 1 = short, 2 = long) */
+	 32,			/* bitsize */
+	 true,			/* pc_relative */
+	 0,			/* bitpos */
+	 complain_overflow_signed, /* complain_on_overflow */
+	 sh_elf_ignore_reloc,	/* special_function */
+	 "R_SH_REL32",		/* name */
 	 false,			/* partial_inplace */
 	 0,			/* src_mask */
 	 0xffffffff,		/* dst_mask */
 	 true),			/* pcrel_offset */
+#endif
 
   /* 8 bit PC relative branch divided by 2.  */
   HOWTO (R_SH_DIR8WPN,		/* type */
@@ -3048,7 +3064,11 @@
 		 section symbol winds up in the output section.  */
 	      sym = local_syms + r_symndx;
 	      if (ELF_ST_TYPE (sym->st_info) == STT_SECTION)
-		goto final_link_relocate;
+		{
+		  r = _bfd_relocate_contents (howto, input_bfd, relocation,
+					      contents + rel->r_offset);
+		  goto relocation_done;
+		}
 
 	      continue;
 	    }
@@ -3251,9 +3271,15 @@
 	      else if (r_type == R_SH_REL32)
 		{
 		  BFD_ASSERT (h != NULL && h->dynindx != -1);
+#if defined(LINUX_ABI)
+		  addend = bfd_get_32 (input_bfd,
+				       (bfd_byte *) contents + rel->r_offset);
+#else
+		  addend = rel->r_addend;
+#endif
 		  relocate = false;
 		  outrel.r_info = ELF32_R_INFO (h->dynindx, R_SH_REL32);
-		  outrel.r_addend = rel->r_addend;
+		  outrel.r_addend = addend;
 		}
 	      else
 		{
@@ -3266,14 +3292,25 @@
 		    {
 		      relocate = true;
 		      outrel.r_info = ELF32_R_INFO (0, R_SH_RELATIVE);
+#if defined(LINUX_ABI)
+		      outrel.r_addend = 0;
+#else
 		      outrel.r_addend = relocation + rel->r_addend;
+#endif
 		    }
 		  else
 		    {
 		      BFD_ASSERT (h->dynindx != -1);
+#if defined(LINUX_ABI)
+		      addend = bfd_get_32 (input_bfd,
+					   ((bfd_byte *) contents
+					    + rel->r_offset));
+#else
+		      addend = rel->r_addend;
+#endif
 		      relocate = false;
 		      outrel.r_info = ELF32_R_INFO (h->dynindx, R_SH_DIR32);
-		      outrel.r_addend = relocation + rel->r_addend;
+		      outrel.r_addend = relocation + addend;
 		    }
 		}
 
@@ -3371,7 +3408,11 @@
 					 + sgot->output_offset
 					 + off);
 		      outrel.r_info = ELF32_R_INFO (0, R_SH_RELATIVE);
+#if defined(LINUX_ABI)
+		      outrel.r_addend = 0;
+#else
 		      outrel.r_addend = relocation;
+#endif
 		      bfd_elf32_swap_reloca_out (output_bfd, &outrel,
 						(((Elf32_External_Rela *)
 						  srelgot->contents)
@@ -3471,6 +3512,7 @@
 	  }
 	}
 
+    relocation_done:
       if (r != bfd_reloc_ok)
 	{
 	  switch (r)
@@ -4237,9 +4279,13 @@
 	  && (h->elf_link_hash_flags & ELF_LINK_HASH_DEF_REGULAR))
 	{
 	  rel.r_info = ELF32_R_INFO (0, R_SH_RELATIVE);
+#if defined(LINUX_ABI)
+	  rel.r_addend = 0;
+#else
 	  rel.r_addend = (h->root.u.def.value
 			  + h->root.u.def.section->output_section->vma
 			  + h->root.u.def.section->output_offset);
+#endif
 	}
       else
 	{
Index: tc-sh.c
===================================================================
RCS file: /cvs/src/src/gas/config/tc-sh.c,v
retrieving revision 1.40
diff -u -r1.40 tc-sh.c
--- tc-sh.c	2001/07/03 15:22:19	1.40
+++ tc-sh.c	2001/09/11 00:39:37
@@ -2834,6 +2834,11 @@
       && fixP->fx_addsy != NULL
       && S_IS_WEAK (fixP->fx_addsy))
     val -= S_GET_VALUE  (fixP->fx_addsy);
+
+  if (OUTPUT_FLAVOR == bfd_target_elf_flavour
+      && fixP->fx_r_type == BFD_RELOC_32_PCREL
+      && fixP->fx_addsy != NULL)
+    val += fixP->fx_where + fixP->fx_frag->fr_address;
 #endif
 
 #ifndef BFD_ASSEMBLER

Re: sh-linux patch for 32-bit relative relocation

[M. R. Brown]

* kaz Kojima <kkojima@rr.iij4u.or.jp> on Tue, Sep 11, 2001:

> Hello,
> 
> Here is a patch for sh-linux to handle the substraction of two
> symbols in different sections correctly. This is essentially same
> with the old patch
>   http://sources.redhat.com/ml/binutils/2001-01/msg00071.html
> which had a possible ABI problem in non linux targets, but is now
> organized as a linux specific patch which doesn't affect non linux
> targets. Formally the change for gas/tc-sh.c affects all sh targets,
> but the effect of this change is ignored by non linux targets since
> src_mask of HOWTO (R_SH_REL32,... is set to 0 in such targets.
> I think that almost all sh-linux developpers are using this patch
> successfully and enough test is done. Without this, we can't get
> working shared libraries for stdc++, X and many other packages.
> 

I can confirm that this patch has been working for me since at least March
of this year.  I've also used this patch to generate generic sh-elf
targets, and I've had no problems with that code.

> Regards,
> 	kaz
> --
> 2001-09-11  Kaz Kojima  <kkojima@rr.iij4u.or.jp>
> 
> 	* bfd/elf32-sh-lin.c (LINUX_ABI): Defined.
> 	* bfd/elf32-sh.c (sh_elf_howto_table): Don't mask source
> 	R_SH_REL32 relocation in linux case.
> 	(sh_elf_relocate_section): Use in-place value in computing
> 	addend for R_SH_REL32 relocation in linux case.
> 

M. R.

Re: sh-linux patch for 32-bit relative relocation

[Hans-Peter Nilsson]

On Tue, 11 Sep 2001, kaz Kojima wrote:


> --- elf32-sh-lin.c      2000/09/07 04:23:43     1.1
> +#define LINUX_ABI 1

> --- elf32-sh.c	2001/05/23 16:38:36	1.29
> +#if defined(LINUX_ABI)

No, this won't do; no #if:s.  But I'll try and make a honest
attempt to fix this properly (without ifdefs, retaining binary
compatibility with your shlibs and hopefully objects) later this
week, using yours and any other test-cases sent to this list.
If you have other test-cases covering bugs in this area that
your previous test-cases do not cover, please send them too;
assembly code and gas+ld options and descriptions of the bugs.
You can send them directly to me if you prefer that, but I'd
prefer them also to be sent to the list.

Does anybody know whether there's any other system that uses ELF
shlibs on SH?  If none does, we can set the shlib relocs
partial-inplace to "false", which is what your patch does (in
part).

Other maintainers with approval rights for SH, are as always,
free to override.

brgds, H-P

Re: sh-linux patch for 32-bit relative relocation

[kaz Kojima]

Hans-Peter Nilsson <hp@bitrange.com> wrote:
>> --- elf32-sh-lin.c      2000/09/07 04:23:43     1.1
>> +#define LINUX_ABI 1
> 
>> --- elf32-sh.c	2001/05/23 16:38:36	1.29
>> +#if defined(LINUX_ABI)
> 
> No, this won't do; no #if:s.  But I'll try and make a honest
> attempt to fix this properly (without ifdefs, retaining binary
> compatibility with your shlibs and hopefully objects) later this
> week, using yours and any other test-cases sent to this list.
> If you have other test-cases covering bugs in this area that
> your previous test-cases do not cover, please send them too;
> assembly code and gas+ld options and descriptions of the bugs.
> You can send them directly to me if you prefer that, but I'd
> prefer them also to be sent to the list.

ok. We'll be very happy if we don't need LINUX_ABI. About SH shlib
relocations, I don't know other bugs and have no new test-cases. 
Thanks,

	kaz