[Bug middle-end/46790] New: [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[Bug middle-end/46790] New: [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jason at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

           Summary: [4.6 regression] EH failures in libstdc++ testsuite
                    with --gc-sections and GNU ld 2.18
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: major
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: jason@gcc.gnu.org
                CC: hubicka@gcc.gnu.org
            Target: x86_64-unknown-linux-gnu


Starting with r167085, test runs on gcc10.fsffrance.org have had a bunch of
failures in the libstdc++ testsuite; the testcases fail with calls to
terminate() because exceptions are not handled properly.  This seems to be due
to .gcc_except_table being thrown away by --gc-sections, which in turn seems to
be due to the change in the name of text subsections; .text.startup.main no
longer matches .gcc_except_table.main, so the heuristic that ld 2.18 uses to
keep .gcc_except_table hunks breaks.

One test that fails is 18_support/uncaught_exception/14026.cc.

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[sje at cup dot hp.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

Steve Ellcey <sje at cup dot hp.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |sje at cup dot hp.com

--- Comment #1 from Steve Ellcey <sje at cup dot hp.com> 2010-12-03 21:19:55 UTC ---
It may be completely unrelated but I have a bunch of failures in C++ on IA64
HP-UX (and probably Linux) involving exceptions that I tracked back to the
latest GNU Assembler rather then any GCC change.  The change was made to the
ToT GAS sources on December 2 and the bug report I submitted is:

http://sourceware.org/bugzilla/show_bug.cgi?id=12287

This report mentions using ld 2.18 but it doesn't say what gas was used,
if it is not ToT then it is unrelated.

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.6.0

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[hubicka at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

Jan Hubicka <hubicka at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2010.12.12 20:59:48
                 CC|                            |hjl.tools at gmail dot com
     Ever Confirmed|0                           |1

--- Comment #2 from Jan Hubicka <hubicka at gcc dot gnu.org> 2010-12-12 20:59:48 UTC ---
I reproduce same issue on gcc14, too.

This seems like previously semi-latent problem. One can definitly declare main
into text.startup by hand. H.J., do you have any idea how to fix/work around
this?  

Steve: HPUX is not using function sections, so it is probably unrelated issue.

Honza

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

--- Comment #3 from H.J. Lu <hjl.tools at gmail dot com> 2010-12-12 21:31:57 UTC ---
There are some gc-sections bug fixes since binutils 2.18:

Fix --gc-sections to detect unresolved symbol in DSO.  PR 11218.
Fix linker --gc-sections with undefined __start_XXX/__stop_XXX symbols.  PR
11133.
Fix linker --gc-sections with SHT_NOTE section.  PR 11143.
Fix "ld -r --gc-sections --entry" crash with COMDAT group.  PR 9727.

You can try the Linux binutils from

http://www.kernel.org/pub/linux/devel/binutils/

to narrow down which binutils fixes it.

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P1

--- Comment #4 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-02-08 13:49:17 UTC ---
So it's really a linker bug.  Can we detect it at configure time and turn
off the fancy renaming?  If we do not fix this bug we should document this
issue in changes.html and adjust the minimum required binutils version
in doc/install.texi (some very ancient versions seem to be mentioned there,
and no target-independent minimal version is suggested).

Marking P1 to force some kind of solution before the release.

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

Steven Bosscher <steven at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |steven at gcc dot gnu.org

--- Comment #5 from Steven Bosscher <steven at gcc dot gnu.org> 2011-02-21 00:10:21 UTC ---
(In reply to comment #4)

Adjusting the minimum required binutils version to what?

For modifying GCC, relatively new versions of libraries and tools are often
required. GMP 4.3.2 for example is from January 2010, autoconf 2.64 was
released in July 2009, and automake 1.11.1 is from December 2009.

But IMHO it should be possible to build and use GCC with relatively old tools.
For example, gmake 3.80 is sufficient, and that's from October 2002. Likewise
for other "Tools/packages necessary for building GCC". GNU binutils 2.18 is
from August 2007, which is not really that old.

But disabling a feature in the compiler depending on a configuration check on
the installed version of a linker it also seems strange to me. You end up with
the same GCC version generating different code depending on the linker.

FWIW, quick link to the revision mentioned to have causes this problem:
http://gcc.gnu.org/viewcvs?view=revision&revision=167085
http://gcc.gnu.org/ml/gcc-patches/2010-11/msg02154.html

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-02-21 00:23:09 UTC ---
So what do you suggest then, if you dislike both updating the minimum version
and disabling the feature based on ld version?  Just document that
--gc-sections will not work with old binutils?

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jason at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

--- Comment #7 from Jason Merrill <jason at gcc dot gnu.org> 2011-02-21 05:30:18 UTC ---
We control lots of things based on whether or not the other tools support a
particular feature.  Doing the same for this issue doesn't seem strange at all
to me.

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-02-21 17:14:10 UTC ---
So, I've tried a few linkers from various RHEL/Fedora distros, and narrowed it
down to the fact that 20071102 ld still fails, while 20080208 ld already works.
There have been pretty big --gc-sections changes in between, starting with:
http://sources.redhat.com/ml/binutils/2007-12/msg00000.html
series and then its follow-ups BZ#5526 and BZ#5604.

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-02-21 17:23:02 UTC ---
Created attachment 23428
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=23428
gcc46-pr46790-configury.patch

Completely untested draft of a configury patch (still no code changes to
actually revert Honza's changes ifdef HAVE_GLD_GC_BUG).  I think writing a
testcase that wouldn't use date strings would be quite entertaining (it would
need to be written in assembler, but in a way that it assembles and does the
right thing on all targets).

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[hubicka at ucw dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

--- Comment #10 from Jan Hubicka <hubicka at ucw dot cz> 2011-02-21 17:30:00 UTC ---
Testing datestamp seems resonable to me. I can do the changes needed to avoid
.text subsections then
(basically the elf implementation should then return NULL)

Honza

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

--- Comment #11 from Steven Bosscher <steven at gcc dot gnu.org> 2011-02-21 17:47:14 UTC ---
(In reply to comment #7)

Interesting, I didn't know that :-)
Do you have an example?

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

--- Comment #12 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-02-21 17:51:00 UTC ---
HAVE_GAS_HIDDEN and HAVE_COMDAT_GROUP checks use ld --version date too.

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #23428|0                           |1
        is obsolete|                            |

--- Comment #13 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-02-21 19:20:32 UTC ---
Created attachment 23429
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=23429
variant 1

So, a few untested or only very lightly tested alternatives.

This one is basically the above configury check plus varasm change to avoid
using .text.startup.*, .text.exit.* etc. with broken linkers.

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

--- Comment #14 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-02-21 19:21:49 UTC ---
Created attachment 23430
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=23430
variant 2

This one will just disable .gcc_except_table.foo for the buggy linkers (in
addition to the even more buggy (and older) ones).

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

--- Comment #15 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-02-21 19:23:36 UTC ---
Created attachment 23431
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=23431
variant 3

This is code-wide like variant 1, but configury is actually checking for the
bug (slightly adapted EH-compatible --gc-section support test can be used for
this).

So, any preferences among these?  Perhaps another alternative would be to use
.gcc_except_table.startup.main and similar, though not sure how hard would that
be.

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

--- Comment #16 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-02-22 10:51:58 UTC ---
Created attachment 23433
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=23433
gcc46-pr46790.patch

The last mentioned variant, which instead of doing configury puts stuff into
.gcc_except_table.startup.main, .gcc_except_table.exit.foo,
.gcc_except_table.unlikely.bar or .gcc_except_table.hot.baz style sections.
Briefly tested with the old and new linker that the testcase is fixed.

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
         AssignedTo|unassigned at gcc dot       |jakub at gcc dot gnu.org
                   |gnu.org                     |

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

--- Comment #17 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-02-28 15:36:41 UTC ---
Author: jakub
Date: Mon Feb 28 15:36:37 2011
New Revision: 170565

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=170565
Log:
    PR middle-end/46790
    * configure.ac (HAVE_LD_EH_GC_SECTIONS_BUG): New test.
    * configure: Regenerated.
    * config.in: Regenerated.
    * varasm.c (default_function_section): Return NULL
    if HAVE_LD_EH_GC_SECTIONS_BUG and decl has implicit
    section name.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config.in
    trunk/gcc/configure
    trunk/gcc/configure.ac
    trunk/gcc/varasm.c

[Bug middle-end/46790] [4.6 regression] EH failures in libstdc++ testsuite with --gc-sections and GNU ld 2.18

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46790

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED

--- Comment #18 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-02-28 16:15:39 UTC ---
Should be fixed now.