A symbol version bug?

A symbol version bug?

[H. J. Lu]

[-- Attachment #1: Type: text/plain, Size: 394 bytes --]

I got

# make
cc -fPIC -c c1.c
cc -fPIC -c c2.c
ld --version-script=c.map -shared -o lib-c1.so c1.o c2.o
objdump --dynamic-sym lib-c1.so| grep bar
00000264 g    DF .text  0000000a  VERS.0      bar
00000258 g    DF .text  0000000a (VERS.0)     bar

I am expecting

...
objdump --dynamic-sym lib-c1.so| grep bar
xxxxxxxx g    DF .text  0000000a (VERS.0)     bar

I will see what I can do.


H.J.

[-- Attachment #2: bug.tar.gz --]
[-- Type: application/x-gzip, Size: 480 bytes --]

Re: A symbol version bug?

[H. J. Lu]

On Fri, Jul 12, 2002 at 08:51:03PM -0700, H. J. Lu wrote:
> I got
> 
> # make
> cc -fPIC -c c1.c
> cc -fPIC -c c2.c
> ld --version-script=c.map -shared -o lib-c1.so c1.o c2.o
> objdump --dynamic-sym lib-c1.so| grep bar
> 00000264 g    DF .text  0000000a  VERS.0      bar
> 00000258 g    DF .text  0000000a (VERS.0)     bar
> 
> I am expecting
> 

Corretion. I am expecting a fatal error.


H.J.

PACTH: Fix duplicated version (Re: A symbol version bug?)

[H. J. Lu]

[-- Attachment #1: Type: text/plain, Size: 891 bytes --]

On Fri, Jul 12, 2002 at 10:43:33PM -0700, H. J. Lu wrote:
> On Fri, Jul 12, 2002 at 08:51:03PM -0700, H. J. Lu wrote:
> > I got
> > 
> > # make
> > cc -fPIC -c c1.c
> > cc -fPIC -c c2.c
> > ld --version-script=c.map -shared -o lib-c1.so c1.o c2.o
> > objdump --dynamic-sym lib-c1.so| grep bar
> > 00000264 g    DF .text  0000000a  VERS.0      bar
> > 00000258 g    DF .text  0000000a (VERS.0)     bar
> > 
> > I am expecting
> > 
> 
> Corretion. I am expecting a fatal error.
> 

Here is the patch. Now I got

/export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new -soname
lib-c.so --version-script=c.map -shared -o lib-c1.so c1.o c2.o
/export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new: foo (c2.o):
duplicated versioned symbol: foo@VERS.0 (c1.o)
/export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new: failed to set
dynamic section sizes: Bad value


H.J.

[-- Attachment #2: bfd-version-dup.patch --]
[-- Type: text/plain, Size: 2016 bytes --]

2002-07-13  H.J. Lu <hjl@gnu.org>

	* elflink.h (elf_link_assign_sym_version): Bail out if a symbol
	with a default version conflicts with a hidden symbol with the
	same version.

--- bfd/elflink.h.dup	Mon Jul  8 08:58:42 2002
+++ bfd/elflink.h	Sat Jul 13 19:48:37 2002
@@ -4345,6 +4345,7 @@ elf_link_assign_sym_version (h, data)
 	    (_("%s: undefined versioned symbol name %s"),
 	     bfd_get_filename (sinfo->output_bfd), h->root.root.string);
 	  bfd_set_error (bfd_error_bad_value);
+error_return:
 	  sinfo->failed = true;
 	  return false;
 	}
@@ -4416,6 +4417,52 @@ elf_link_assign_sym_version (h, data)
 	      (*bed->elf_backend_hide_symbol) (info, h, true);
 	    }
 	}
+
+      /* We need to check if there is a conflict.  */
+      if (h->dynindx != -1)
+	{
+	  const char *verstr, *name;
+	  size_t namelen, verlen, newlen;
+	  char *newname, *p;
+	  struct elf_link_hash_entry *newh;
+
+	  name = h->root.root.string;
+	  namelen = strlen (name);
+	  verstr = h->verinfo.vertree->name;
+	  verlen = strlen (verstr);
+	  newlen = namelen + verlen + 2;
+
+	  newname = (char *) bfd_malloc ((bfd_size_type) newlen);
+	  if (newname == NULL)
+	    {
+	      bfd_set_error (bfd_error_no_memory);
+	      goto error_return;
+	    }
+	  memcpy (newname, name, namelen);
+
+	  /* Check the hidden symbol.  */
+	  p = newname + namelen;
+	  *p++ = ELF_VER_CHR;
+	  memcpy (p, verstr, verlen + 1);
+	  newh = elf_link_hash_lookup (elf_hash_table (info), newname,
+				    false, false, false);
+
+	  if (newh && newh->root.type == bfd_link_hash_defined)
+	    {
+	      /* We find a versioned definition. Bail out.  */
+	      (*_bfd_error_handler)
+		(_("%s (%s): duplicated versioned symbol: %s (%s)"),
+		 h->root.root.string, 
+		 bfd_get_filename (h->root.u.def.section->owner),
+		 newname,
+		 bfd_get_filename (newh->root.u.def.section->owner));
+	      free (newname);
+	      bfd_set_error (bfd_error_bad_value);
+	      goto error_return;
+	    }
+
+	  free (newname);
+	}
     }
 
   return true;

Re: PACTH: Fix duplicated version (Re: A symbol version bug?)

[H. J. Lu]

[-- Attachment #1: Type: text/plain, Size: 2114 bytes --]

On Sat, Jul 13, 2002 at 07:56:44PM -0700, H. J. Lu wrote:
> On Fri, Jul 12, 2002 at 10:43:33PM -0700, H. J. Lu wrote:
> > On Fri, Jul 12, 2002 at 08:51:03PM -0700, H. J. Lu wrote:
> > > I got
> > > 
> > > # make
> > > cc -fPIC -c c1.c
> > > cc -fPIC -c c2.c
> > > ld --version-script=c.map -shared -o lib-c1.so c1.o c2.o
> > > objdump --dynamic-sym lib-c1.so| grep bar
> > > 00000264 g    DF .text  0000000a  VERS.0      bar
> > > 00000258 g    DF .text  0000000a (VERS.0)     bar
> > > 
> > > I am expecting
> > > 
> > 
> > Corretion. I am expecting a fatal error.
> > 
> 
> Here is the patch. Now I got
> 
> /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new -soname
> lib-c.so --version-script=c.map -shared -o lib-c1.so c1.o c2.o
> /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new: foo (c2.o):
> duplicated versioned symbol: foo@VERS.0 (c1.o)
> /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new: failed to set
> dynamic section sizes: Bad value
> 

Second thought. I think it should work. Here is the new patch and a
testcase. Now I got:

# make
cc    -c -o foo.o foo.c
cc    -c -o c0.o c0.c
cc -Wl,-soname,lib-c.so -shared -o lib-c.so c0.o
cc -o foo -Wl,-rpath,. foo.o lib-c.so
cc -fPIC -c c1.c
cc -fPIC -c c2.c
/export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new -soname
lib-c.so --version-script=c.map -shared -o lib-c1.so c1.o c2.o
/export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new -soname lib-c.so --version-script=c.map -shared -o lib-c2.so c2.o c1.o
objdump --dynamic-sym lib-c1.so | grep VER
00000280 g    DF .text  00000021 (VERS.0)     foo
00000000 g    DO *ABS*  00000000  VERS.0      VERS.0
0000025c g    DF .text  00000021 (VERS.0)     bar
objdump --dynamic-sym lib-c2.so | grep VER
0000028c g    DF .text  00000021 (VERS.0)     foo
00000000 g    DO *ABS*  00000000  VERS.0      VERS.0
00000268 g    DF .text  00000021 (VERS.0)     bar
./foo
bar: 0
foo: 0
cp lib-c1.so lib-c.so
./foo
bar: 1
foo: 2
cp lib-c2.so lib-c.so
./foo
bar: 1
foo: 2

Jack, please try this to see if gcc 2.95.4 works without any glibc
changes.


H.J.

[-- Attachment #2: bug.tar.gz --]
[-- Type: application/x-gzip, Size: 798 bytes --]

[-- Attachment #3: bfd-version-dup.patch --]
[-- Type: text/plain, Size: 1770 bytes --]

2002-07-14  H.J. Lu <hjl@gnu.org>

	* elflink.h (elf_link_assign_sym_version): Hide the default
	definition if there is a hidden versioned definition.

--- bfd/elflink.h.dup	Mon Jul  8 08:58:42 2002
+++ bfd/elflink.h	Sun Jul 14 06:49:57 2002
@@ -4345,6 +4345,7 @@ elf_link_assign_sym_version (h, data)
 	    (_("%s: undefined versioned symbol name %s"),
 	     bfd_get_filename (sinfo->output_bfd), h->root.root.string);
 	  bfd_set_error (bfd_error_bad_value);
+error_return:
 	  sinfo->failed = true;
 	  return false;
 	}
@@ -4416,6 +4417,44 @@ elf_link_assign_sym_version (h, data)
 	      (*bed->elf_backend_hide_symbol) (info, h, true);
 	    }
 	}
+
+      /* We need to check if a hidden versioned definition should
+	 override the default one.  */
+      if (h->dynindx != -1)
+	{
+	  const char *verstr, *name;
+	  size_t namelen, verlen, newlen;
+	  char *newname;
+	  struct elf_link_hash_entry *newh;
+
+	  name = h->root.root.string;
+	  namelen = strlen (name);
+	  verstr = h->verinfo.vertree->name;
+	  verlen = strlen (verstr);
+	  newlen = namelen + verlen + 2;
+
+	  newname = (char *) bfd_malloc ((bfd_size_type) newlen);
+	  if (newname == NULL)
+	    {
+	      bfd_set_error (bfd_error_no_memory);
+	      goto error_return;
+	    }
+	  memcpy (newname, name, namelen);
+
+	  /* Check the hidden versioned definition.  */
+	  p = newname + namelen;
+	  *p++ = ELF_VER_CHR;
+	  memcpy (p, verstr, verlen + 1);
+	  newh = elf_link_hash_lookup (elf_hash_table (info), newname,
+				    false, false, false);
+
+	  if (newh && newh->root.type == bfd_link_hash_defined)
+	    /* We find a versioned definition. Hide the default
+	       one.  */
+	    (*bed->elf_backend_hide_symbol) (info, h, true);
+
+	  free (newname);
+	}
     }
 
   return true;

Re: PACTH: Fix duplicated version (Re: A symbol version bug?)

[H. J. Lu]

[-- Attachment #1: Type: text/plain, Size: 2343 bytes --]

On Sun, Jul 14, 2002 at 06:56:20AM -0700, H. J. Lu wrote:
> On Sat, Jul 13, 2002 at 07:56:44PM -0700, H. J. Lu wrote:
> > On Fri, Jul 12, 2002 at 10:43:33PM -0700, H. J. Lu wrote:
> > > On Fri, Jul 12, 2002 at 08:51:03PM -0700, H. J. Lu wrote:
> > > > I got
> > > > 
> > > > # make
> > > > cc -fPIC -c c1.c
> > > > cc -fPIC -c c2.c
> > > > ld --version-script=c.map -shared -o lib-c1.so c1.o c2.o
> > > > objdump --dynamic-sym lib-c1.so| grep bar
> > > > 00000264 g    DF .text  0000000a  VERS.0      bar
> > > > 00000258 g    DF .text  0000000a (VERS.0)     bar
> > > > 
> > > > I am expecting
> > > > 
> > > 
> > > Corretion. I am expecting a fatal error.
> > > 
> > 
> > Here is the patch. Now I got
> > 
> > /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new -soname
> > lib-c.so --version-script=c.map -shared -o lib-c1.so c1.o c2.o
> > /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new: foo (c2.o):
> > duplicated versioned symbol: foo@VERS.0 (c1.o)
> > /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new: failed to set
> > dynamic section sizes: Bad value
> > 
> 
> Second thought. I think it should work. Here is the new patch and a
> testcase. Now I got:
> 
> # make
> cc    -c -o foo.o foo.c
> cc    -c -o c0.o c0.c
> cc -Wl,-soname,lib-c.so -shared -o lib-c.so c0.o
> cc -o foo -Wl,-rpath,. foo.o lib-c.so
> cc -fPIC -c c1.c
> cc -fPIC -c c2.c
> /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new -soname
> lib-c.so --version-script=c.map -shared -o lib-c1.so c1.o c2.o
> /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new -soname lib-c.so --version-script=c.map -shared -o lib-c2.so c2.o c1.o
> objdump --dynamic-sym lib-c1.so | grep VER
> 00000280 g    DF .text  00000021 (VERS.0)     foo
> 00000000 g    DO *ABS*  00000000  VERS.0      VERS.0
> 0000025c g    DF .text  00000021 (VERS.0)     bar
> objdump --dynamic-sym lib-c2.so | grep VER
> 0000028c g    DF .text  00000021 (VERS.0)     foo
> 00000000 g    DO *ABS*  00000000  VERS.0      VERS.0
> 00000268 g    DF .text  00000021 (VERS.0)     bar
> ./foo
> bar: 0
> foo: 0
> cp lib-c1.so lib-c.so
> ./foo
> bar: 1
> foo: 2
> cp lib-c2.so lib-c.so
> ./foo
> bar: 1
> foo: 2
> 
> Jack, please try this to see if gcc 2.95.4 works without any glibc
> changes.
> 

Ooops. Wrong one. Please use this one.


H.J.

[-- Attachment #2: bfd-version-dup.patch --]
[-- Type: text/plain, Size: 1800 bytes --]

2002-07-14  H.J. Lu <hjl@gnu.org>

	* elflink.h (elf_link_assign_sym_version): Hide the default
	definition if there is a hidden versioned definition.

--- bfd/elflink.h.dup	Mon Jul  8 08:58:42 2002
+++ bfd/elflink.h	Sun Jul 14 07:06:31 2002
@@ -4345,6 +4345,7 @@ elf_link_assign_sym_version (h, data)
 	    (_("%s: undefined versioned symbol name %s"),
 	     bfd_get_filename (sinfo->output_bfd), h->root.root.string);
 	  bfd_set_error (bfd_error_bad_value);
+error_return:
 	  sinfo->failed = true;
 	  return false;
 	}
@@ -4416,6 +4417,44 @@ elf_link_assign_sym_version (h, data)
 	      (*bed->elf_backend_hide_symbol) (info, h, true);
 	    }
 	}
+
+      /* We need to check if a hidden versioned definition should
+	 override the default one.  */
+      if (h->dynindx != -1 && h->verinfo.vertree != NULL)
+	{
+	  const char *verstr, *name;
+	  size_t namelen, verlen, newlen;
+	  char *newname;
+	  struct elf_link_hash_entry *newh;
+
+	  name = h->root.root.string;
+	  namelen = strlen (name);
+	  verstr = h->verinfo.vertree->name;
+	  verlen = strlen (verstr);
+	  newlen = namelen + verlen + 2;
+
+	  newname = (char *) bfd_malloc ((bfd_size_type) newlen);
+	  if (newname == NULL)
+	    {
+	      bfd_set_error (bfd_error_no_memory);
+	      goto error_return;
+	    }
+	  memcpy (newname, name, namelen);
+
+	  /* Check the hidden versioned definition.  */
+	  p = newname + namelen;
+	  *p++ = ELF_VER_CHR;
+	  memcpy (p, verstr, verlen + 1);
+	  newh = elf_link_hash_lookup (elf_hash_table (info), newname,
+				    false, false, false);
+
+	  if (newh && newh->root.type == bfd_link_hash_defined)
+	    /* We find a versioned definition. Hide the default
+	       one.  */
+	    (*bed->elf_backend_hide_symbol) (info, h, true);
+
+	  free (newname);
+	}
     }
 
   return true;

Re: PACTH: Fix duplicated version (Re: A symbol version bug?)

[H. J. Lu]

[-- Attachment #1: Type: text/plain, Size: 2604 bytes --]

On Sun, Jul 14, 2002 at 07:09:53AM -0700, H. J. Lu wrote:
> On Sun, Jul 14, 2002 at 06:56:20AM -0700, H. J. Lu wrote:
> > On Sat, Jul 13, 2002 at 07:56:44PM -0700, H. J. Lu wrote:
> > > On Fri, Jul 12, 2002 at 10:43:33PM -0700, H. J. Lu wrote:
> > > > On Fri, Jul 12, 2002 at 08:51:03PM -0700, H. J. Lu wrote:
> > > > > I got
> > > > > 
> > > > > # make
> > > > > cc -fPIC -c c1.c
> > > > > cc -fPIC -c c2.c
> > > > > ld --version-script=c.map -shared -o lib-c1.so c1.o c2.o
> > > > > objdump --dynamic-sym lib-c1.so| grep bar
> > > > > 00000264 g    DF .text  0000000a  VERS.0      bar
> > > > > 00000258 g    DF .text  0000000a (VERS.0)     bar
> > > > > 
> > > > > I am expecting
> > > > > 
> > > > 
> > > > Corretion. I am expecting a fatal error.
> > > > 
> > > 
> > > Here is the patch. Now I got
> > > 
> > > /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new -soname
> > > lib-c.so --version-script=c.map -shared -o lib-c1.so c1.o c2.o
> > > /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new: foo (c2.o):
> > > duplicated versioned symbol: foo@VERS.0 (c1.o)
> > > /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new: failed to set
> > > dynamic section sizes: Bad value
> > > 
> > 
> > Second thought. I think it should work. Here is the new patch and a
> > testcase. Now I got:
> > 
> > # make
> > cc    -c -o foo.o foo.c
> > cc    -c -o c0.o c0.c
> > cc -Wl,-soname,lib-c.so -shared -o lib-c.so c0.o
> > cc -o foo -Wl,-rpath,. foo.o lib-c.so
> > cc -fPIC -c c1.c
> > cc -fPIC -c c2.c
> > /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new -soname
> > lib-c.so --version-script=c.map -shared -o lib-c1.so c1.o c2.o
> > /export/build/gnu/binutils-debug/build-full-i686-linux/ld/ld-new -soname lib-c.so --version-script=c.map -shared -o lib-c2.so c2.o c1.o
> > objdump --dynamic-sym lib-c1.so | grep VER
> > 00000280 g    DF .text  00000021 (VERS.0)     foo
> > 00000000 g    DO *ABS*  00000000  VERS.0      VERS.0
> > 0000025c g    DF .text  00000021 (VERS.0)     bar
> > objdump --dynamic-sym lib-c2.so | grep VER
> > 0000028c g    DF .text  00000021 (VERS.0)     foo
> > 00000000 g    DO *ABS*  00000000  VERS.0      VERS.0
> > 00000268 g    DF .text  00000021 (VERS.0)     bar
> > ./foo
> > bar: 0
> > foo: 0
> > cp lib-c1.so lib-c.so
> > ./foo
> > bar: 1
> > foo: 2
> > cp lib-c2.so lib-c.so
> > ./foo
> > bar: 1
> > foo: 2
> > 
> > Jack, please try this to see if gcc 2.95.4 works without any glibc
> > changes.
> > 
> 
> Ooops. Wrong one. Please use this one.
> 
> 

Ooops. I missed the weak definition. Here is the new patch.


H.J.

[-- Attachment #2: bfd-version-dup.patch --]
[-- Type: text/plain, Size: 1862 bytes --]

2002-07-14  H.J. Lu <hjl@gnu.org>

	* elflink.h (elf_link_assign_sym_version): Hide the default
	definition if there is a hidden versioned definition.

--- bfd/elflink.h.dup	Mon Jul  8 08:58:42 2002
+++ bfd/elflink.h	Sun Jul 14 07:41:59 2002
@@ -4345,6 +4345,7 @@ elf_link_assign_sym_version (h, data)
 	    (_("%s: undefined versioned symbol name %s"),
 	     bfd_get_filename (sinfo->output_bfd), h->root.root.string);
 	  bfd_set_error (bfd_error_bad_value);
+error_return:
 	  sinfo->failed = true;
 	  return false;
 	}
@@ -4416,6 +4417,46 @@ elf_link_assign_sym_version (h, data)
 	      (*bed->elf_backend_hide_symbol) (info, h, true);
 	    }
 	}
+
+      /* We need to check if a hidden versioned definition should
+	 hide the default one.  */
+      if (h->dynindx != -1 && h->verinfo.vertree != NULL)
+	{
+	  const char *verstr, *name;
+	  size_t namelen, verlen, newlen;
+	  char *newname;
+	  struct elf_link_hash_entry *newh;
+
+	  name = h->root.root.string;
+	  namelen = strlen (name);
+	  verstr = h->verinfo.vertree->name;
+	  verlen = strlen (verstr);
+	  newlen = namelen + verlen + 2;
+
+	  newname = (char *) bfd_malloc ((bfd_size_type) newlen);
+	  if (newname == NULL)
+	    {
+	      bfd_set_error (bfd_error_no_memory);
+	      goto error_return;
+	    }
+	  memcpy (newname, name, namelen);
+
+	  /* Check the hidden versioned definition.  */
+	  p = newname + namelen;
+	  *p++ = ELF_VER_CHR;
+	  memcpy (p, verstr, verlen + 1);
+	  newh = elf_link_hash_lookup (elf_hash_table (info), newname,
+				    false, false, false);
+
+	  if (newh
+	      && (newh->root.type == bfd_link_hash_defined
+		  || newh->root.type == bfd_link_hash_defweak))
+	    /* We find a hidden versioned definition. Hide the default
+	       one.  */
+	    (*bed->elf_backend_hide_symbol) (info, h, true);
+
+	  free (newname);
+	}
     }
 
   return true;

Re: PACTH: Fix duplicated version (Re: A symbol version bug?)

[Jack Howarth]

HJ,
   Using this last version of the patch, applied to binutils
2.12.90.0.14, I was able to build a binutils which passed all
of make check. Using this new binutils, I was able to build
debian's current glibc 2.2.5-9 package with the libgcc-compat
patch added in. There were no link problems with libc6.so this
time with the patched binutils and the resulting glibc passes all
of make check. Looks like this might be the fix we need.
                       Jack

Re: PACTH: Fix duplicated version (Re: A symbol version bug?)

[Alan Modra]

On Sun, Jul 14, 2002 at 06:56:20AM -0700, H. J. Lu wrote:
> On Sat, Jul 13, 2002 at 07:56:44PM -0700, H. J. Lu wrote:
> > On Fri, Jul 12, 2002 at 10:43:33PM -0700, H. J. Lu wrote:
> > > On Fri, Jul 12, 2002 at 08:51:03PM -0700, H. J. Lu wrote:
> > > > I am expecting
> > > Corretion. I am expecting a fatal error.
> Second thought. I think it should work.

That's quite a few changes of mind. :)  Can you tell us why you "think
it should work"?  Why doesn't bar defined by
__asm__(".symver _old_bar,bar@VERS.0");
not conflict with another bar, set to VERS.0 by a version script?

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Re: PACTH: Fix duplicated version (Re: A symbol version bug?)

[H. J. Lu]

On Mon, Jul 15, 2002 at 01:08:09AM +0930, Alan Modra wrote:
> On Sun, Jul 14, 2002 at 06:56:20AM -0700, H. J. Lu wrote:
> > On Sat, Jul 13, 2002 at 07:56:44PM -0700, H. J. Lu wrote:
> > > On Fri, Jul 12, 2002 at 10:43:33PM -0700, H. J. Lu wrote:
> > > > On Fri, Jul 12, 2002 at 08:51:03PM -0700, H. J. Lu wrote:
> > > > > I am expecting
> > > > Corretion. I am expecting a fatal error.
> > Second thought. I think it should work.
> 
> That's quite a few changes of mind. :)  Can you tell us why you "think
> it should work"?  Why doesn't bar defined by
> __asm__(".symver _old_bar,bar@VERS.0");
> not conflict with another bar, set to VERS.0 by a version script?

That is because by using

__asm__(".symver _old_bar,bar@VERS.0");

the intention is the definition of bar with version VERS.0 is _old_bar,
The version script only specifies there is a bar with version VERS.0,
nothing more. As my testcase show, it has its use.


H.J.

Re: PACTH: Fix duplicated version (Re: A symbol version bug?)

[Alan Modra]

On Sun, Jul 14, 2002 at 04:32:52PM -0700, H. J. Lu wrote:
> On Mon, Jul 15, 2002 at 01:08:09AM +0930, Alan Modra wrote:
> > That's quite a few changes of mind. :)  Can you tell us why you "think
> > it should work"?  Why doesn't bar defined by
> > __asm__(".symver _old_bar,bar@VERS.0");
> > not conflict with another bar, set to VERS.0 by a version script?
> 
> That is because by using
> 
> __asm__(".symver _old_bar,bar@VERS.0");
> 
> the intention is the definition of bar with version VERS.0 is _old_bar,
> The version script only specifies there is a bar with version VERS.0,
> nothing more. As my testcase show, it has its use.

I suppose it's fair enough to say that a .symver takes precedence over
a version script.  Would you please make a note of this in ld.texinfo?
Your patch is OK, but remove the unnecessary bfd_set_error on malloc
failure.  bfd_malloc will do this for you.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Re: PACTH: Fix duplicated version (Re: A symbol version bug?)

[H. J. Lu]

[-- Attachment #1: Type: text/plain, Size: 1062 bytes --]

On Mon, Jul 15, 2002 at 10:17:44AM +0930, Alan Modra wrote:
> On Sun, Jul 14, 2002 at 04:32:52PM -0700, H. J. Lu wrote:
> > On Mon, Jul 15, 2002 at 01:08:09AM +0930, Alan Modra wrote:
> > > That's quite a few changes of mind. :)  Can you tell us why you "think
> > > it should work"?  Why doesn't bar defined by
> > > __asm__(".symver _old_bar,bar@VERS.0");
> > > not conflict with another bar, set to VERS.0 by a version script?
> > 
> > That is because by using
> > 
> > __asm__(".symver _old_bar,bar@VERS.0");
> > 
> > the intention is the definition of bar with version VERS.0 is _old_bar,
> > The version script only specifies there is a bar with version VERS.0,
> > nothing more. As my testcase show, it has its use.
> 
> I suppose it's fair enough to say that a .symver takes precedence over
> a version script.  Would you please make a note of this in ld.texinfo?
> Your patch is OK, but remove the unnecessary bfd_set_error on malloc
> failure.  bfd_malloc will do this for you.

Ok. Here is the updated one. I will check it in shortly.

Thanks.


H.J.

[-- Attachment #2: bfd-version-dup.patch --]
[-- Type: text/plain, Size: 2610 bytes --]

bfd/
2002-07-14  H.J. Lu <hjl@gnu.org>

	* elflink.h (elf_link_assign_sym_version): Hide the default
	definition if there is a hidden versioned definition.

ld/
2002-07-14  H.J. Lu <hjl@gnu.org>

	* ld.texinfo: Document a .symver takes precedence over a
	version script.

--- binutils/bfd/elflink.h.dup	Mon Jul  8 08:58:42 2002
+++ binutils/bfd/elflink.h	Sun Jul 14 18:25:09 2002
@@ -4345,6 +4345,7 @@ elf_link_assign_sym_version (h, data)
 	    (_("%s: undefined versioned symbol name %s"),
 	     bfd_get_filename (sinfo->output_bfd), h->root.root.string);
 	  bfd_set_error (bfd_error_bad_value);
+error_return:
 	  sinfo->failed = true;
 	  return false;
 	}
@@ -4416,6 +4417,43 @@ elf_link_assign_sym_version (h, data)
 	      (*bed->elf_backend_hide_symbol) (info, h, true);
 	    }
 	}
+
+      /* We need to check if a hidden versioned definition should
+	 hide the default one.  */
+      if (h->dynindx != -1 && h->verinfo.vertree != NULL)
+	{
+	  const char *verstr, *name;
+	  size_t namelen, verlen, newlen;
+	  char *newname;
+	  struct elf_link_hash_entry *newh;
+
+	  name = h->root.root.string;
+	  namelen = strlen (name);
+	  verstr = h->verinfo.vertree->name;
+	  verlen = strlen (verstr);
+	  newlen = namelen + verlen + 2;
+
+	  newname = (char *) bfd_malloc ((bfd_size_type) newlen);
+	  if (newname == NULL)
+	    goto error_return;
+	  memcpy (newname, name, namelen);
+
+	  /* Check the hidden versioned definition.  */
+	  p = newname + namelen;
+	  *p++ = ELF_VER_CHR;
+	  memcpy (p, verstr, verlen + 1);
+	  newh = elf_link_hash_lookup (elf_hash_table (info), newname,
+				    false, false, false);
+
+	  if (newh
+	      && (newh->root.type == bfd_link_hash_defined
+		  || newh->root.type == bfd_link_hash_defweak))
+	    /* We find a hidden versioned definition. Hide the default
+	       one.  */
+	    (*bed->elf_backend_hide_symbol) (info, h, true);
+
+	  free (newname);
+	}
     }
 
   return true;
--- binutils/ld/ld.texinfo.dup	Sat Jun 22 21:49:03 2002
+++ binutils/ld/ld.texinfo	Sun Jul 14 18:36:02 2002
@@ -3762,7 +3762,8 @@ __asm__(".symver original_foo,foo@@VERS_
 in the C source file.  This renames the function @samp{original_foo} to
 be an alias for @samp{foo} bound to the version node @samp{VERS_1.1}.
 The @samp{local:} directive can be used to prevent the symbol
-@samp{original_foo} from being exported.
+@samp{original_foo} from being exported. A @samp{.symver} directive
+takes precedence over a version script.
 
 The second GNU extension is to allow multiple versions of the same
 function to appear in a given shared library.  In this way you can make