bfd is broken

bfd is broken

[H . J . Lu]

This patch

http://sources.redhat.com/ml/binutils/2002-01/msg00094.html

breaks bfd which causes "make check" to fild in ld:

Running /home/hjl/work/gnu/src/binutils/binutils/ld/testsuite/ld-bootstrap/bootstrap.exp
...
FAIL: bootstrap
FAIL: bootstrap with strip
FAIL: bootstrap with --static
FAIL: bootstrap with --traditional-format
FAIL: bootstrap with --no-keep-memory

I am using gcc 2.96-99 from Red Hat and binutils is configured with

# ../configure --enable-64-bit-bfd --enable-targets=all

on Linux/x86.



H.J.

Re: bfd is broken

[H . J . Lu]

On Thu, Jan 10, 2002 at 06:27:30PM -0800, H . J . Lu wrote:
> This patch
> 
> http://sources.redhat.com/ml/binutils/2002-01/msg00094.html
> 
> breaks bfd which causes "make check" to fild in ld:
> 
> Running /home/hjl/work/gnu/src/binutils/binutils/ld/testsuite/ld-bootstrap/bootstrap.exp
> ...
> FAIL: bootstrap
> FAIL: bootstrap with strip
> FAIL: bootstrap with --static
> FAIL: bootstrap with --traditional-format
> FAIL: bootstrap with --no-keep-memory
> 
> I am using gcc 2.96-99 from Red Hat and binutils is configured with
> 
> # ../configure --enable-64-bit-bfd --enable-targets=all
> 
> on Linux/x86.

Hi Geoffrey,

If you can show me how to reproduce it, I will look into it. A .o file
should be fine.



H.J.

Re: bfd is broken

[Geoff Keating]

The configuration I used to test the original patch was x86 Red Hat
Linux 7 with all updates applied, using the Red Hat GNUPro 00r1
compiler, without special configure flags; and the bootstrap test
passed.  When I got your message I ran a build with the toplevel
binutils CVS tree, the latest bundled compiler gcc-2.96-85, and your two
special configure flags; and the bootstrap test still passed.  In fact
all tests pass or xfail.

I wouldn't expect SEC_EXCLUDE to be used much when building ld.  How
did you determine that it was my patch?

Can you be more specific about your system?

-- 
- Geoffrey Keating <geoffk@geoffk.org> <geoffk@redhat.com>

Re: bfd is broken

[H . J . Lu]

On Thu, Jan 10, 2002 at 08:54:51PM -0800, Geoff Keating wrote:
> 
> The configuration I used to test the original patch was x86 Red Hat
> Linux 7 with all updates applied, using the Red Hat GNUPro 00r1
> compiler, without special configure flags; and the bootstrap test
> passed.  When I got your message I ran a build with the toplevel
> binutils CVS tree, the latest bundled compiler gcc-2.96-85, and your two
> special configure flags; and the bootstrap test still passed.  In fact
> all tests pass or xfail.
> 
> I wouldn't expect SEC_EXCLUDE to be used much when building ld.  How
> did you determine that it was my patch?

Backing out your patch fixes ld. Can you tell me how to reproduce your
bug? The .c/.o/.s files are fine.

> 
> Can you be more specific about your system?
> 

Some string constances have bogus values. The only differences are I
am using gcc 2.96-99 and binutils 2.11.92.0.12.3. 


H.J.

Re: bfd is broken

[Jakub Jelinek]

On Thu, Jan 10, 2002 at 11:03:13PM -0800, H . J . Lu wrote:
> On Thu, Jan 10, 2002 at 08:54:51PM -0800, Geoff Keating wrote:
> > 
> > The configuration I used to test the original patch was x86 Red Hat
> > Linux 7 with all updates applied, using the Red Hat GNUPro 00r1
> > compiler, without special configure flags; and the bootstrap test
> > passed.  When I got your message I ran a build with the toplevel
> > binutils CVS tree, the latest bundled compiler gcc-2.96-85, and your two
> > special configure flags; and the bootstrap test still passed.  In fact
> > all tests pass or xfail.
> > 
> > I wouldn't expect SEC_EXCLUDE to be used much when building ld.  How
> > did you determine that it was my patch?
> 
> Backing out your patch fixes ld. Can you tell me how to reproduce your
> bug? The .c/.o/.s files are fine.

I'd really like to see testcase, but looking at Geoff patch I see what can
be wrong: _bfd_merged_section_offset needs to be called even if
(isec->flags & SEC_EXCLUDE), provided SEC_EXCLUDE has been added because of
_cooked_size 0.

E.g. if a.s has:
.section ".rodata.str1.1", "aMS", @progbits, 1
.LC1: .asciiz "foobar"
.LC2: .asciiz "bogusstr"

and b.s has:
.section ".rodata.str1.1", "aMS", @progbits, 1
.LC3: .asciiz "bar"
.LC4: .asciiz "str"

then b.o's .rodata.str1.1 section will be SEC_EXCLUDE, but .LC3 needs to be
translated to .LC1 + 3 and .LC4 to .LC2 + 5.

	Jakub

Re: bfd is broken

[H . J . Lu]

On Fri, Jan 11, 2002 at 10:59:10AM +0100, Jakub Jelinek wrote:
> On Thu, Jan 10, 2002 at 11:03:13PM -0800, H . J . Lu wrote:
> > On Thu, Jan 10, 2002 at 08:54:51PM -0800, Geoff Keating wrote:
> > > 
> > > The configuration I used to test the original patch was x86 Red Hat
> > > Linux 7 with all updates applied, using the Red Hat GNUPro 00r1
> > > compiler, without special configure flags; and the bootstrap test
> > > passed.  When I got your message I ran a build with the toplevel
> > > binutils CVS tree, the latest bundled compiler gcc-2.96-85, and your two
> > > special configure flags; and the bootstrap test still passed.  In fact
> > > all tests pass or xfail.
> > > 
> > > I wouldn't expect SEC_EXCLUDE to be used much when building ld.  How
> > > did you determine that it was my patch?
> > 
> > Backing out your patch fixes ld. Can you tell me how to reproduce your
> > bug? The .c/.o/.s files are fine.
> 
> I'd really like to see testcase, but looking at Geoff patch I see what can
> be wrong: _bfd_merged_section_offset needs to be called even if
> (isec->flags & SEC_EXCLUDE), provided SEC_EXCLUDE has been added because of
> _cooked_size 0.
> 
> E.g. if a.s has:
> .section ".rodata.str1.1", "aMS", @progbits, 1
> .LC1: .asciiz "foobar"
> .LC2: .asciiz "bogusstr"
> 
> and b.s has:
> .section ".rodata.str1.1", "aMS", @progbits, 1
> .LC3: .asciiz "bar"
> .LC4: .asciiz "str"
> 
> then b.o's .rodata.str1.1 section will be SEC_EXCLUDE, but .LC3 needs to be
> translated to .LC1 + 3 and .LC4 to .LC2 + 5.

I believe it is what happened in my case. May I suggest to back out
the broken patch? We will fix it later when a testcase is provided?


H.J.

Re: bfd is broken

[Jakub Jelinek]

On Fri, Jan 11, 2002 at 09:22:22AM -0800, H . J . Lu wrote:
> > E.g. if a.s has:
> > .section ".rodata.str1.1", "aMS", @progbits, 1
> > .LC1: .asciiz "foobar"
> > .LC2: .asciiz "bogusstr"
> > 
> > and b.s has:
> > .section ".rodata.str1.1", "aMS", @progbits, 1
> > .LC3: .asciiz "bar"
> > .LC4: .asciiz "str"
> > 
> > then b.o's .rodata.str1.1 section will be SEC_EXCLUDE, but .LC3 needs to be
> > translated to .LC1 + 3 and .LC4 to .LC2 + 5.
> 
> I believe it is what happened in my case. May I suggest to back out
> the broken patch? We will fix it later when a testcase is provided?

Geoff sais where is the testcase (gcc testsuite gcsec-1.c).
I was looking for a testcase where it failed for you, so if you think the
above is what happens, no further need for it (though it should probably
be added as a testcase for ld).

	Jakub

Re: bfd is broken

[H . J . Lu]

On Fri, Jan 11, 2002 at 12:24:51PM -0500, Jakub Jelinek wrote:
> On Fri, Jan 11, 2002 at 09:22:22AM -0800, H . J . Lu wrote:
> > > E.g. if a.s has:
> > > .section ".rodata.str1.1", "aMS", @progbits, 1
> > > .LC1: .asciiz "foobar"
> > > .LC2: .asciiz "bogusstr"
> > > 
> > > and b.s has:
> > > .section ".rodata.str1.1", "aMS", @progbits, 1
> > > .LC3: .asciiz "bar"
> > > .LC4: .asciiz "str"
> > > 
> > > then b.o's .rodata.str1.1 section will be SEC_EXCLUDE, but .LC3 needs to be
> > > translated to .LC1 + 3 and .LC4 to .LC2 + 5.
> > 
> > I believe it is what happened in my case. May I suggest to back out
> > the broken patch? We will fix it later when a testcase is provided?
> 
> Geoff sais where is the testcase (gcc testsuite gcsec-1.c).

I tried. I couldn't get gcsec-1.c to fail on Linux/x86. I am wondering
how it failed for him.

> I was looking for a testcase where it failed for you, so if you think the
> above is what happens, no further need for it (though it should probably
> be added as a testcase for ld).

I will see what I can do. I may write one which only runs on Linux/x86.
It should be good enough for catching breakage.


H.J.

Re: bfd is broken

[Jakub Jelinek]

On Fri, Jan 11, 2002 at 09:34:15AM -0800, H . J . Lu wrote:
> I tried. I couldn't get gcsec-1.c to fail on Linux/x86. I am wondering
> how it failed for him.

Executing on host: /usr/src/gcc/obj/gcc/xgcc -B/usr/src/gcc/obj/gcc/ /usr/src/gcc/gcc/testsuite/gcc.dg/special/gcsec-1.c  -ffunction-sections -fdata-sections -Wl,--gc-sections -static   -lm   -o ./a.out    (timeout = 300)
collect2: ld terminated with signal 11 [Segmentation fault], core dumped
compiler exited with status 1

> > I was looking for a testcase where it failed for you, so if you think the
> > above is what happens, no further need for it (though it should probably
> > be added as a testcase for ld).
> 
> I will see what I can do. I may write one which only runs on Linux/x86.
> It should be good enough for catching breakage.

Well, there are 3 ways how to write it.
Either write a plain C testcase which will test SHF_MERGE only if gcc
supports it, or gcc -S and sed the .s file to make plain .rodata sections
.section ".rodata.str", "aMS", @progbits, 1
or have testcase fully in assembly (in which case it will be for one arch
only).

	Jakub

Re: bfd is broken

[H . J . Lu]

On Fri, Jan 11, 2002 at 12:38:45PM -0500, Jakub Jelinek wrote:
> On Fri, Jan 11, 2002 at 09:34:15AM -0800, H . J . Lu wrote:
> > I tried. I couldn't get gcsec-1.c to fail on Linux/x86. I am wondering
> > how it failed for him.
> 
> Executing on host: /usr/src/gcc/obj/gcc/xgcc -B/usr/src/gcc/obj/gcc/ /usr/src/gcc/gcc/testsuite/gcc.dg/special/gcsec-1.c  -ffunction-sections -fdata-sections -Wl,--gc-sections -static   -lm   -o ./a.out    (timeout = 300)
> collect2: ld terminated with signal 11 [Segmentation fault], core dumped
> compiler exited with status 1

Thanks. I am enclosing a patch.

> 
> > > I was looking for a testcase where it failed for you, so if you think the
> > > above is what happens, no further need for it (though it should probably
> > > be added as a testcase for ld).
> > 
> > I will see what I can do. I may write one which only runs on Linux/x86.
> > It should be good enough for catching breakage.
> 
> Well, there are 3 ways how to write it.
> Either write a plain C testcase which will test SHF_MERGE only if gcc
> supports it, or gcc -S and sed the .s file to make plain .rodata sections
> .section ".rodata.str", "aMS", @progbits, 1
> or have testcase fully in assembly (in which case it will be for one arch
> only).
> 

I don't see there are any testcases for merge. Let me think about it.


H.J.
---
2002-01-11  H.J. Lu <hjl@gnu.org>

	* elflink.h (elf_link_input_bfd): Revert the change made on
	2002-01-07. Don't ask for the merged offset only if sec_info
	is NULL instead.

--- bfd/elflink.h.merge	Thu Jan 10 17:48:45 2002
+++ bfd/elflink.h	Fri Jan 11 09:49:27 2002
@@ -6375,8 +6375,7 @@ elf_link_input_bfd (finfo, input_bfd)
 	  isec = section_from_elf_index (input_bfd, isym->st_shndx);
 	  if (isec
 	      && elf_section_data (isec)->sec_info_type == ELF_INFO_TYPE_MERGE
-	      && (finfo->info->relocateable
-		  || ! (isec->flags & SEC_EXCLUDE))
+	      && elf_section_data (isec)->sec_info != NULL
 	      && ELF_ST_TYPE (isym->st_info) != STT_SECTION)
 	    isym->st_value =
 	      _bfd_merged_section_offset (output_bfd, &isec,

Re: bfd is broken

[Geoff Keating]

> Date: Fri, 11 Jan 2002 10:59:10 +0100
> From: Jakub Jelinek <jakub@redhat.com>
> Cc: Geoff Keating <geoffk@redhat.com>, binutils@sources.redhat.com
> Reply-To: Jakub Jelinek <jakub@redhat.com>
> 
> On Thu, Jan 10, 2002 at 11:03:13PM -0800, H . J . Lu wrote:
> > On Thu, Jan 10, 2002 at 08:54:51PM -0800, Geoff Keating wrote:
> > > 
> > > The configuration I used to test the original patch was x86 Red Hat
> > > Linux 7 with all updates applied, using the Red Hat GNUPro 00r1
> > > compiler, without special configure flags; and the bootstrap test
> > > passed.  When I got your message I ran a build with the toplevel
> > > binutils CVS tree, the latest bundled compiler gcc-2.96-85, and your two
> > > special configure flags; and the bootstrap test still passed.  In fact
> > > all tests pass or xfail.
> > > 
> > > I wouldn't expect SEC_EXCLUDE to be used much when building ld.  How
> > > did you determine that it was my patch?
> > 
> > Backing out your patch fixes ld. Can you tell me how to reproduce your
> > bug? The .c/.o/.s files are fine.
> 
> I'd really like to see testcase, but looking at Geoff patch I see what can
> be wrong: _bfd_merged_section_offset needs to be called even if
> (isec->flags & SEC_EXCLUDE), provided SEC_EXCLUDE has been added because of
> _cooked_size 0.
> 
> E.g. if a.s has:
> .section ".rodata.str1.1", "aMS", @progbits, 1
> .LC1: .asciiz "foobar"
> .LC2: .asciiz "bogusstr"
> 
> and b.s has:
> .section ".rodata.str1.1", "aMS", @progbits, 1
> .LC3: .asciiz "bar"
> .LC4: .asciiz "str"
> 
> then b.o's .rodata.str1.1 section will be SEC_EXCLUDE, but .LC3 needs to be
> translated to .LC1 + 3 and .LC4 to .LC2 + 5.

That makes sense.  Why do we set SEC_EXCLUDE on this section, if it
has size 0 anyway and so wouldn't take up any space in the output?

-- 
- Geoffrey Keating <geoffk@geoffk.org> <geoffk@redhat.com>

Re: bfd is broken

[H . J . Lu]

On Fri, Jan 11, 2002 at 12:38:45PM -0500, Jakub Jelinek wrote:
> On Fri, Jan 11, 2002 at 09:34:15AM -0800, H . J . Lu wrote:
> > I tried. I couldn't get gcsec-1.c to fail on Linux/x86. I am wondering
> > how it failed for him.
> 
> Executing on host: /usr/src/gcc/obj/gcc/xgcc -B/usr/src/gcc/obj/gcc/ /usr/src/gcc/gcc/testsuite/gcc.dg/special/gcsec-1.c  -ffunction-sections -fdata-sections -Wl,--gc-sections -static   -lm   -o ./a.out    (timeout = 300)
> collect2: ld terminated with signal 11 [Segmentation fault], core dumped
> compiler exited with status 1
> 
> > > I was looking for a testcase where it failed for you, so if you think the
> > > above is what happens, no further need for it (though it should probably
> > > be added as a testcase for ld).
> > 
> > I will see what I can do. I may write one which only runs on Linux/x86.
> > It should be good enough for catching breakage.
> 
> Well, there are 3 ways how to write it.
> Either write a plain C testcase which will test SHF_MERGE only if gcc
> supports it, or gcc -S and sed the .s file to make plain .rodata sections
> .section ".rodata.str", "aMS", @progbits, 1
> or have testcase fully in assembly (in which case it will be for one arch
> only).

I have a testcase in assembly. But it is quite tricky. You have to
use as from 2.11.92.0.12.3 to see the failure. I don't know how useful
the testcase is.


H.J.

Re: bfd is broken

[Jakub Jelinek]

On Fri, Jan 11, 2002 at 11:13:58AM -0800, Geoff Keating wrote:
> That makes sense.  Why do we set SEC_EXCLUDE on this section, if it
> has size 0 anyway and so wouldn't take up any space in the output?

Because bfd ATM is not able to set _cooked_size = 0 (_cooked_size == 0
means use _raw_size). It really needs to be changed soon, but it will mean
changing a lot of code.

	Jakub

[PATCH] Fix GC and SHF_MERGE interaction (was Re: bfd is broken)

[Jakub Jelinek]

On Fri, Jan 11, 2002 at 11:13:58AM -0800, Geoff Keating wrote:
> That makes sense.  Why do we set SEC_EXCLUDE on this section, if it
> has size 0 anyway and so wouldn't take up any space in the output?

Here is a fix:

It worked before I changed ->stab_info, ->merge_info into ->sec_info_type,
->sec_merge (_bfd_merge_sections was clearing merge_info if it encountered
something (e.g. GC) deleted one of the SEC_MERGE sections), but now it just
cleared it and didn't know there is section info type encoded somewhere.
One solution would be to always check both that
sec_info_type == ELF_INFO_TYPE_MERGE && sec_info != NULL
but IMHO just clearing sec_info_type in this case is better.
Bootstrapped on i386-redhat-linux, no failures, fixes gcsec-1.c.
Ok to commit?

2002-01-14  Jakub Jelinek  <jakub@redhat.com>

	* elflink.h (elf_link_input_bfd): Back out 2002-01-07 change.
	* elf.c (merge_sections_remove_hook): New function.
	(_bfd_elf_merge_sections): Pass it as 3rd argument to
	_bfd_merge_sections.
	* libbfd-in.h (_bfd_merge_sections): Add 3rd argument.
	* libbfd.h: Rebuilt.
	* merge.c (_bfd_merge_sections): Add remove_hook argument.
	Call remove_hook if a SEC_EXCLUDE section is encountered.

--- bfd/elflink.h.jj	Mon Jan 14 17:02:47 2002
+++ bfd/elflink.h	Mon Jan 14 17:44:31 2002
@@ -6375,8 +6375,6 @@ elf_link_input_bfd (finfo, input_bfd)
 	  isec = section_from_elf_index (input_bfd, isym->st_shndx);
 	  if (isec
 	      && elf_section_data (isec)->sec_info_type == ELF_INFO_TYPE_MERGE
-	      && (finfo->info->relocateable
-		  || ! (isec->flags & SEC_EXCLUDE))
 	      && ELF_ST_TYPE (isym->st_info) != STT_SECTION)
 	    isym->st_value =
 	      _bfd_merged_section_offset (output_bfd, &isec,
--- bfd/elf.c.jj	Mon Jan 14 17:02:46 2002
+++ bfd/elf.c	Mon Jan 14 18:17:59 2002
@@ -53,6 +53,7 @@ static boolean swap_out_syms PARAMS ((bf
 static boolean copy_private_bfd_data PARAMS ((bfd *, bfd *));
 static char *elf_read PARAMS ((bfd *, file_ptr, bfd_size_type));
 static boolean setup_group PARAMS ((bfd *, Elf_Internal_Shdr *, asection *));
+static void merge_sections_remove_hook PARAMS ((bfd *, asection *));
 static void elf_fake_sections PARAMS ((bfd *, asection *, PTR));
 static void set_group_contents PARAMS ((bfd *, asection *, PTR));
 static boolean assign_section_numbers PARAMS ((bfd *));
@@ -770,6 +771,20 @@ bfd_elf_generic_reloc (abfd,
   return bfd_reloc_continue;
 }
 \f
+/* Make sure sec_info_type is cleared if sec_info is cleared too.  */
+
+static void
+merge_sections_remove_hook (abfd, sec)
+     bfd *abfd ATTRIBUTE_UNUSED;
+     asection *sec;
+{
+  struct bfd_elf_section_data *sec_data;
+    
+  sec_data = elf_section_data (sec);
+  BFD_ASSERT (sec_data->sec_info_type == ELF_INFO_TYPE_MERGE);
+  sec_data->sec_info_type = ELF_INFO_TYPE_NONE;
+}
+
 /* Finish SHF_MERGE section merging.  */
 
 boolean
@@ -780,7 +795,8 @@ _bfd_elf_merge_sections (abfd, info)
   if (!is_elf_hash_table (info))
     return false;
   if (elf_hash_table (info)->merge_info)
-    _bfd_merge_sections (abfd, elf_hash_table (info)->merge_info);
+    _bfd_merge_sections (abfd, elf_hash_table (info)->merge_info,
+			 merge_sections_remove_hook);
   return true;
 }
 \f
--- bfd/libbfd-in.h.jj	Mon Jan 14 17:02:48 2002
+++ bfd/libbfd-in.h	Mon Jan 14 17:50:16 2002
@@ -478,7 +478,7 @@ extern boolean _bfd_merge_section
 /* Attempt to merge SEC_MERGE sections.  */
 
 extern boolean _bfd_merge_sections
-  PARAMS ((bfd *, PTR));
+  PARAMS ((bfd *, PTR, void (*)(bfd *, asection *)));
 
 /* Write out a merged section.  */
 
--- bfd/libbfd.h.jj	Mon Jan 14 17:02:48 2002
+++ bfd/libbfd.h	Mon Jan 14 17:50:28 2002
@@ -483,7 +483,7 @@ extern boolean _bfd_merge_section
 /* Attempt to merge SEC_MERGE sections.  */
 
 extern boolean _bfd_merge_sections
-  PARAMS ((bfd *, PTR));
+  PARAMS ((bfd *, PTR, void (*)(bfd *, asection *)));
 
 /* Write out a merged section.  */
 
--- bfd/merge.c.jj	Fri Oct  5 13:02:24 2001
+++ bfd/merge.c	Mon Jan 14 17:53:10 2002
@@ -771,9 +771,10 @@ alloc_failure:
    with _bfd_merge_section.  */
 
 boolean
-_bfd_merge_sections (abfd, xsinfo)
+_bfd_merge_sections (abfd, xsinfo, remove_hook)
      bfd *abfd ATTRIBUTE_UNUSED;
      PTR xsinfo;
+     void (*remove_hook) PARAMS((bfd *, asection *));
 {
   struct sec_merge_info *sinfo;
 
@@ -792,7 +793,11 @@ _bfd_merge_sections (abfd, xsinfo)
       /* Record the sections into the hash table.  */
       for (secinfo = sinfo->chain; secinfo; secinfo = secinfo->next)
 	if (secinfo->sec->flags & SEC_EXCLUDE)
-	  *secinfo->psecinfo = NULL;
+	  {
+	    *secinfo->psecinfo = NULL;
+	    if (remove_hook)
+	      (*remove_hook) (abfd, secinfo->sec);
+	  }
 	else if (! record_section (sinfo, secinfo))
 	  break;
 


	Jakub

Re: [PATCH] Fix GC and SHF_MERGE interaction (was Re: bfd is broken)

[Alan Modra]

On Mon, Jan 14, 2002 at 05:23:06PM +0100, Jakub Jelinek wrote:
> 
> 	* elflink.h (elf_link_input_bfd): Back out 2002-01-07 change.
> 	* elf.c (merge_sections_remove_hook): New function.
> 	(_bfd_elf_merge_sections): Pass it as 3rd argument to
> 	_bfd_merge_sections.
> 	* libbfd-in.h (_bfd_merge_sections): Add 3rd argument.
> 	* libbfd.h: Rebuilt.
> 	* merge.c (_bfd_merge_sections): Add remove_hook argument.
> 	Call remove_hook if a SEC_EXCLUDE section is encountered.

OK.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre