[Bug rtl-optimization/46114] New: [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[Bug rtl-optimization/46114] New: [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[ro at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

           Summary: [4.6 regression] g++ SEGV when built with gld on
                    Solaris 10+/x86
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: ro@gcc.gnu.org
                CC: bernds@codesourcery.com
              Host: i386-pc-solaris2.1[01]
            Target: i386-pc-solaris2.1[01]
             Build: i386-pc-solaris2.1[01]


When I recently tried to bootstrap mainline on Solaris 11 (and later 10)/x86
with CVS gas and gld, many C++ testcases failed.  It turned out that g++
SEGVs like this:

% g++ gcc.o
Segmentation Fault

It crashes like this:

Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 1 (LWP 1)]
generate_option (opt_index=902, arg=0x8074535 <error reading variable>, 
    value=1, lang_mask=65536, decoded=0x1)
    at /var/gcc/reghunt/trunk/gcc/opts-common.c:862
862       decoded->arg = arg;
(gdb) where
#0  generate_option (opt_index=902, arg=0x8074535 <error reading variable>, 
    value=1, lang_mask=65536, decoded=0x1)
    at /var/gcc/reghunt/trunk/gcc/opts-common.c:862
#1  0x0805a9cb in lang_specific_driver (in_decoded_options=0x80474d8, 
    in_decoded_options_count=0x80474dc, in_added_libraries=0x8093744)
    at /var/gcc/reghunt/trunk/gcc/cp/g++spec.c:324
#2  0x0804c764 in process_command (decoded_options_count=2, 
    decoded_options=0x8093ff8) at /var/gcc/reghunt/trunk/gcc/gcc.c:3597
#3  0x08052f9a in main (argc=2, argv=0x8047664)
    at /var/gcc/reghunt/trunk/gcc/gcc.c:6214

i.e. decoded is invalid.  After searching around for a binutils change that
caused this, it turned out that the problem only happens when gas and gld
are in use, but also with gas/gld 2.20.1.  A gcc reghunt identified the
following patch as the culprit:

2010-09-23  Bernd Schmidt  <bernds@codesourcery.com>

        PR rtl-optimization/44374
        * basic-block.h (enum bb_flags): Add BB_MODIFIED.
        * df-core.c (df_set_bb_dirty): Set it.
        * ifcvt.c (find_memory): Remove function.
        (dead_or_predicable): Use can_move_insns_across.
        * df.h (can_move_insns_across): Declare function.
        * cfgcleanup.c (block_was_dirty): New static variable.
        (try_crossjump_bb, try_forward_edges): Test BB_MODIFIED flag rather
        than df_get_bb_dirty.
        (try_head_merge_bb): New static function.
        (try_optimize_cfg): Call it.  Call df_analyze if block_was_dirty
        is set.
        * df-problems.c: Include "target.h"
        (df_simulate_find_uses): New static function.
        (MEMREF_NORMAL, MEMREF_VOLATILE): New macros.
        (find_memory, find_memory_store): New static functions.
        (can_move_insns_across): New function.
        * Makefile.in (df-problems.o): Update dependencies.

I've not yet identified the mis-compilation caused.

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[ro at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

--- Comment #1 from Rainer Orth <ro at gcc dot gnu.org> 2010-10-21 20:31:57 UTC ---
Created attachment 22109
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22109
main program of testcase

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[ro at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

--- Comment #2 from Rainer Orth <ro at gcc dot gnu.org> 2010-10-21 20:32:47 UTC ---
Created attachment 22110
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22110
miscompiled part of the test

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[ro at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

Rainer Orth <ro at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2010.10.21 20:36:21
     Ever Confirmed|0                           |1

--- Comment #3 from Rainer Orth <ro at gcc dot gnu.org> 2010-10-21 20:36:21 UTC ---
I could reproduce the bug with the attached testcase (go.c with main,
pr46114.c reduced from g++spec.c).

Compile with

% gcc -O2 -o go go.c pr46114.c

After the culprit patch, the program prints

opt_index = 902 arg = 8049253 value = 1 lang_mask = 65536 decoded = 1
opt_index = 902 arg = 8049264 value = 1 lang_mask = 65536 decoded = 804a90c
opt_index = 1019 arg = 0 value = 1 lang_mask = 65536 decoded = 804a938

i.e. decoded (a pointer) is invalid.  Before, one gets

opt_index = 902 arg = 8049273 value = 1 lang_mask = 65536 decoded = 804a900
opt_index = 902 arg = 8049284 value = 1 lang_mask = 65536 decoded = 804a92c
opt_index = 1019 arg = 0 value = 1 lang_mask = 65536 decoded = 804a958

The bug is extremely sensitive to the details of the input, so I didn't try
to reduce it further.

Please fix, this is really ugly and makes gcc with gld completely useless.

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hjl.tools at gmail dot com

--- Comment #4 from H.J. Lu <hjl.tools at gmail dot com> 2010-10-21 20:41:06 UTC ---
Can you try

http://gcc.gnu.org/ml/gcc-patches/2010-10/msg01858.html

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[ro at CeBiTec dot Uni-Bielefeld.DE]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

--- Comment #5 from ro at CeBiTec dot Uni-Bielefeld.DE <ro at CeBiTec dot Uni-Bielefeld.DE> 2010-10-21 20:57:42 UTC ---
> --- Comment #4 from H.J. Lu <hjl.tools at gmail dot com> 2010-10-21 20:41:06 UTC ---
> Can you try
>
> http://gcc.gnu.org/ml/gcc-patches/2010-10/msg01858.html

Doesn't make a difference unfortunately.

Thanks.
        Rainer

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[bernds at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

--- Comment #6 from Bernd Schmidt <bernds at gcc dot gnu.org> 2010-10-21 23:33:47 UTC ---
I'm assuming you run the testcase on Solaris? Can you provide good/bad assembly
output?

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[ro at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

--- Comment #7 from Rainer Orth <ro at gcc dot gnu.org> 2010-10-22 10:09:26 UTC ---
Created attachment 22113
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22113
working assembler output

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[ro at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

--- Comment #8 from Rainer Orth <ro at gcc dot gnu.org> 2010-10-22 10:10:13 UTC ---
Created attachment 22114
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22114
miscompiled assembler output

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[ro at CeBiTec dot Uni-Bielefeld.DE]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

--- Comment #9 from ro at CeBiTec dot Uni-Bielefeld.DE <ro at CeBiTec dot Uni-Bielefeld.DE> 2010-10-22 10:15:20 UTC ---
> --- Comment #6 from Bernd Schmidt <bernds at gcc dot gnu.org> 2010-10-21 23:33:47 UTC ---
> I'm assuming you run the testcase on Solaris? Can you provide good/bad assembly
> output?

Sure, both attached.  Each of them was compiled with gcc -g -O2 -S.

As could be expected, the bug has nothing to do with wether or nor gld
is in use, only with the details of the input.  It can be reproduced
with a gcc configured to use Sun as and ld instead of gas and gld.  The
attached -S output is for a compiler configured to use CVS gas and gld,
though.  The bad one is from r164552 (your patch), the good one from
r164551 (the last working rev).

Thanks.
    Rainer

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[bernds at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

Bernd Schmidt <bernds at gcc dot gnu.org> changed:

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

--- Comment #10 from Bernd Schmidt <bernds at gcc dot gnu.org> 2010-10-22 13:24:53 UTC ---
I'll also need preprocessed source since I'm lacking Solaris includes.  I'll be
away for a few weeks so there might be no progress on this bug during that
time.

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[ro at CeBiTec dot Uni-Bielefeld.DE]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

--- Comment #11 from ro at CeBiTec dot Uni-Bielefeld.DE <ro at CeBiTec dot Uni-Bielefeld.DE> 2010-10-22 13:35:08 UTC ---
> --- Comment #10 from Bernd Schmidt <bernds at gcc dot gnu.org> 2010-10-22 13:24:53 UTC ---
> I'll also need preprocessed source since I'm lacking Solaris includes.  I'll be
> away for a few weeks so there might be no progress on this bug during that
> time.

It should suffice to replace the includes with declarations of size_t,
abort, calloc, malloc, strcmp, and strlen in pr46114.c:

typedef unsigned int size_t;
extern void abort(void);
extern void *calloc(size_t, size_t);
extern void *malloc(size_t);
extern int strcmp(const char *, const char *);
extern size_t strlen(const char *);

I've replaced those by the corresponding includes so one can compile and
run the testcase on other platforms.

I'll be away until next tuesday myself and won't be reading email during
that time.  I very much hope this can be fixed before 4.6.0 is released.

    Rainer

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

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

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

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |WAITING

--- Comment #12 from H.J. Lu <hjl.tools at gmail dot com> 2010-11-03 17:10:07 UTC ---
Please try revision 166259.

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[ro at CeBiTec dot Uni-Bielefeld.DE]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

--- Comment #13 from ro at CeBiTec dot Uni-Bielefeld.DE <ro at CeBiTec dot Uni-Bielefeld.DE> 2010-11-04 18:48:46 UTC ---
> --- Comment #12 from H.J. Lu <hjl.tools at gmail dot com> 2010-11-03 17:10:07 UTC ---
> Please try revision 166259.

As expected, the reversion fixes the testcase.  A full bootstrap with
CVS gas and gld is currently running, and g++ is already running
successfully.

    Rainer

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |RESOLVED
                 CC|                            |steven at gcc dot gnu.org
         Resolution|                            |FIXED

--- Comment #14 from Steven Bosscher <steven at gcc dot gnu.org> 2010-11-08 07:30:47 UTC ---
Fixed by the commit that reverted bernds' patch.

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[bernds at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

--- Comment #15 from Bernd Schmidt <bernds at gcc dot gnu.org> 2010-11-25 16:05:36 UTC ---
Can you go back to the failing sources and test whether
  http://gcc.gnu.org/ml/gcc-patches/2010-10/msg02569.html
fixes the problem?

For future bug reports, please use gcc -v -save-temps, then post the
preprocessed source and the exact options that were passed to cc1. I've managed
to reproduce the code generation, but only with an explicit "-march=i686" which
required some guessing.

[Bug rtl-optimization/46114] [4.6 regression] g++ SEGV when built with gld on Solaris 10+/x86

[ro at CeBiTec dot Uni-Bielefeld.DE]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46114

--- Comment #16 from ro at CeBiTec dot Uni-Bielefeld.DE <ro at CeBiTec dot Uni-Bielefeld.DE> 2010-11-25 16:25:35 UTC ---
> --- Comment #15 from Bernd Schmidt <bernds at gcc dot gnu.org> 2010-11-25 16:05:36 UTC ---
> Can you go back to the failing sources and test whether
>   http://gcc.gnu.org/ml/gcc-patches/2010-10/msg02569.html
> fixes the problem?

Looks good indeed.  I didn't try a full bootstrap but just up to the
point where g++ failed before.

> For future bug reports, please use gcc -v -save-temps, then post the
> preprocessed source and the exact options that were passed to cc1. I've managed
> to reproduce the code generation, but only with an explicit "-march=i686" which
> required some guessing.

Ok, will do.  Thanks for fixing this.

    Rainer