[Bug c++/14113] New: -O2 corrupts stack for small example

[Bug c++/14113] New: -O2 corrupts stack for small example

[gcc at cohi dot at]

SYMPTOM

The following code example crashes in the sense that it acts as if it couldn't
find the (obviously present) try block.

SNIPPET

The following function seems to be the critical one, the full code of the
program is included below. As shown, the problem occurs when compiled with -W
-Wall -O2 and goes away if either -O2 is omitted, or one of the two marked
places is changed:

void A::t2( const std::string & s1, const std::string & s2 )
{
   int i;
   if ( ! get( i ) )
   {
         do { std::ostringstream oss; oss << __FILE__ << __LINE__ << " " <<
"kasimir"; throw oss.str(); } while ( 1 );  // normally a THROW() macro

         // Change to "throw std::string( "kasimir" );" lets the problem
disappear, althouth this is _not_ the throw that will be executed.
   }
   switch ( i )
   {
      case 1:
         f1( s1, s2 );
         break;
      case 2:
         f2( s1, s2, false );
         break;
      case 3:
         f2( s1, s2, true );
         break;
// Change case 2 and case 3 to
//    case 2:
//    case 3:
//       f2( s1, s2, i == 3 );
//       break;
// makes the problem disappear.
      case 7:
      case 8:
         f3( s1, s2 );
         break;
   }
}

COMPILER

gcc-3.3-20040209/bin/g++ -v
Reading specs from
/opt/global/gcc-3.3-20040209/lib/gcc-lib/sparc-sun-solaris2.8/3.3.3/specs
Configured with: ./configure --prefix=/opt/global/gcc-3.3-20040209
--enable-shared --enable-threads --with-cpu=v9 --with-gnu-as
--with-as=/opt/global/gcc-3.3-32-solaris8-108528-20/bin/as
--enable-languages=c,c++ --with-gnu-ld
--with-ld=/opt/global/gcc-3.3-32-solaris8-108528-20/bin/ld --disable-multilib
Thread model: posix
gcc version 3.3.3 20040209 (prerelease)

gcc-3.3-32-solaris8-108528-20/bin/g++ -v 
Reading specs from
gcc-3.3-32-solaris8-108528-20/bin/../lib/gcc-lib/sparc-sun-solaris2.8/3.3/specs
Configured with: ../gcc-3.3/configure
--prefix=/opt/global/gcc-3.3-32-solaris8-108528-20 --enable-shared
--enable-threads --with-cpu=v9 --with-gnu-as
--with-as=/opt/global/gcc-3.3-32-solaris8-108528-20/bin/as --with-gnu-ld
--with-ld=/opt/global/gcc-3.3-32-solaris8-108528-20/bin/ld --disable-multilib
Thread model: posix
gcc version 3.3

gcc-3.3.1-DEPRECATED/bin/g++ -v
Reading specs from
gcc-3.3.1-DEPRECATED/bin/../lib/gcc-lib/sparc-sun-solaris2.8/3.3.1/specs
Configured with: ./configure
--prefix=/opt/global/gcc-3.3.1-32-solaris8-108528-20 --enable-shared
--enable-threads --with-cpu=v9 --with-gnu-as
--with-as=/opt/global/gcc-3.3-32-solaris8-108528-20/bin/as
--enable-languages=c,c++ --with-gnu-ld
--with-ld=/opt/global/gcc-3.3-32-solaris8-108528-20/bin/ld --disable-multilib
Thread model: posix
gcc version 3.3.1

(The 3.3.1 is modified - it contains the patch for bug 12301.)

SOURCE

The source is put into two files to prevent inlining...

bug02.hh:

#include <string>
#include <sstream>

bool get( int & );

struct A
{
   void t2( const std::string &, const std::string & );

   void f1( const std::string &, const std::string & );
   void f2( const std::string & s1, const std::string & s2, bool );
   void f3( const std::string &, const std::string & );
};

#define THROW( MeSSaGe ) \
do { std::ostringstream oss; oss << __FILE__ << __LINE__ << ' ' << MeSSaGe;
throw oss.str(); } while( 1 )

bug02-broken.cc:

#include <string>
#include <iostream>

#include "bug02.hh"

void A::t2( const std::string & s1, const std::string & s2 )
{
   int i;
   if ( ! get( i ) )
   {
      THROW( "kasimir" );
   }
   switch ( i )
   {
      case 1:
         f1( s1, s2 );
         break;
      case 2:
         f2( s1, s2, false );
         break;
      case 3:
         f2( s1, s2, true );
         break;
      case 7:
      case 8:
         f3( s1, s2 );
         break;
   }
}

int main( int, char ** )
{
   A * a = new A;

   try
   {
      a->t2( std::string( "foo" ), std::string( "bar" ) );
   }
   catch ( std::string & s )
   {
      std::cerr << s << '\n';
   }
   return 0;
}

bug02-lib.cc:

#include "bug02.hh"

bool get( int & i )
{
   i = 3;
   return true;
}


void A::f1( const std::string &, const std::string & )
{
}

void A::f2( const std::string & s1, const std::string & s2, bool )
{
   // At this point it doesn't seem to matter how the exception is thrown.
   // This is the exception that is thrown, and that is not caught by the
   // catch in main() - unless one of the three changes outlined above is
   // made.

        throw std::string( "nasenbaer" );
//   THROW( s1 << s2 );
}

void A::f3( const std::string &, const std::string & )
{
}

STACK

An excerpt from the stacktrace obtained from our full software that has a
crash/signal/assert/uncaught-exception handler that executes pstack to get a
stacktrace of the crash - seems that something is broken so that it doesn't find
the catch block for the exception.

 fe1f0d4c _thrp_kill (0, 1, 6, fe20c000, 1, fe21321c) + f8
 fe34ba74 raise    (6, 6, ffbedc88, 42c2, 6, 2c0e5c) + 40
 fe33591c abort    (fe3bc000, 0, ffbedd28, 189d1c, 0, 0) + c0
 00189d2c _ZN4Base14terminate_funcEv (1edf90, ffbeddd4, 23f768, 2c0e50, 0, 1) + 8c
 feee59f4 _ZN10__cxxabiv111__terminateEPFvvE (189ca0, 1be1dc, ffbee2a8, 0, 1, 0) + 4
 feee58d4 __gxx_personality_v0 (3, 1cfab4, 474e5543, 432b2b00, a98bb8, ffbee2a8)
+ 4ac
 fe3e5460 _Unwind_RaiseException_Phase2 (2, ffbee2a8, 0, ffbee5f8, 4, 0) + 7c
 fe3e578c _Unwind_Resume (a98be8, ffffffff, 1ed400, 68, 5fba, 1bc008) + 50

-- 
           Summary: -O2 corrupts stack for small example
           Product: gcc
           Version: 3.3
            Status: UNCONFIRMED
          Severity: critical
          Priority: P2
         Component: c++
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: gcc at cohi dot at
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: sparc-sun-solaris2.8
  GCC host triplet: sparc-sun-solaris2.8
GCC target triplet: sparc-sun-solaris2.8


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113

[Bug middle-end/14113] -O2 corrupts stack for small example

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-02-11 17:53 -------
Does this happen with -fno-delayed-branch?

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|c++                         |middle-end
           Keywords|                            |wrong-code
            Version|3.3                         |3.3.3


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113

[Bug optimization/14113] [3.3 regression] -O2 corrupts stack

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2004-02-11 20:47 -------
Confirmed with GCC 3.3.3pre, a regression from GCC 2.95.3.  Works in GCC 3.4pre.
 And -fno-delayed-branch doesn't help.




-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ebotcazou at gcc dot gnu dot
                   |                            |org
             Status|UNCONFIRMED                 |NEW
          Component|middle-end                  |optimization
     Ever Confirmed|                            |1
      Known to fail|                            |3.2.2 3.2.3 3.3.2 3.3.3
      Known to work|                            |2.95.3
   Last reconfirmed|0000-00-00 00:00:00         |2004-02-11 20:47:05
               date|                            |
            Summary|-O2 corrupts stack for small|[3.3 regression] -O2
                   |example                     |corrupts stack
   Target Milestone|---                         |3.3.4


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113

[Bug target/14113] [3.3 regression] -O2 corrupts stack

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2004-02-11 21:55 -------
The culprits are the last 2 remaining define_peepholes of the SPARC port, that I
have already scrapped for GCC 3.4.  They change the value of the PC return
register without updating the CFIs, so the unwind mechanism is fooled.

The workaround is to disable them with -fno-peephole.  They are already disabled
by default on UltraSPARC because they hurt performance on modern CPUs.

I'll try to get the fix integrated in 3.3.3 but I don't promise anything.


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot gnu   |ebotcazou at gcc dot gnu dot
                   |dot org                     |org
             Status|NEW                         |ASSIGNED
          Component|optimization                |target


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113

[Bug target/14113] [3.3 regression] -O2 corrupts stack

[gdr at integrable-solutions dot net]

------- Additional Comments From gdr at integrable-solutions dot net  2004-02-11 22:34 -------
Subject: Re:  [3.3 regression] -O2 corrupts stack

"ebotcazou at gcc dot gnu dot org" <gcc-bugzilla@gcc.gnu.org> writes:

| I'll try to get the fix integrated in 3.3.3 but I don't promise anything.

I suppose you meant 3.3.4 :-)

-- Gaby


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113

[Bug target/14113] [3.3/3.4 regression] -O2 corrupts stack

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2004-02-12 06:59 -------
Correction: the peepholes have been scrapped on mainline, but not on 3.4 branch
so we have a latent regression there too.

Andrew, you would have been right with -fno-delayed-branch if the option was not
itself broken :-)  This is PR target/10567.


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[3.3 regression] -O2        |[3.3/3.4 regression] -O2
                   |corrupts stack              |corrupts stack


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113

[Bug target/14113] [3.3/3.4 regression] -O2 corrupts stack

[gcc at cohi dot at]

------- Additional Comments From gcc at cohi dot at  2004-02-12 08:44 -------
(In reply to comment #3)

> The workaround is to disable them with -fno-peephole.  They are already disabled
> by default on UltraSPARC because they hurt performance on modern CPUs.

The workaround works - thank you all (again!) for the quick reply - I wouldn't
want to put our software into production with this problem still around...

PS: What exactly does "disabled by default on Ultra" mean? We are running on
UltraSPARC III but still triggered this problem. Is this a GCC bulid option or
do we need the correct compile options for our software to make sure the correct
specs are used? For now we patch the Makefile to use -fno-peephole.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113

[Bug target/14113] [3.3/3.4 regression] -O2 corrupts stack

[gdr at integrable-solutions dot net]

------- Additional Comments From gdr at integrable-solutions dot net  2004-02-12 09:04 -------
Subject: Re:  [3.3/3.4 regression] -O2 corrupts stack

"ebotcazou at gcc dot gnu dot org" <gcc-bugzilla@gcc.gnu.org> writes:

| Correction: the peepholes have been scrapped on mainline, but not on
| 3.4 branch so we have a latent regression there too.

So, are you going to apply your patch or not?  Your patch is the only
one I'm waiting for right now...

-- Gaby


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113

[Bug target/14113] [3.3/3.4 regression] -O2 corrupts stack

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2004-02-12 10:11 -------
Yes, I'll apply it as soon as testing is complete.  As of this writing, 3 of the
5 SPARC machines have completed bootstrap and are regtesting.  ETA: 4 hours.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113

[Bug target/14113] [3.3/3.4 regression] -O2 corrupts stack

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2004-02-12 10:23 -------
(In reply to comment #6)

> PS: What exactly does "disabled by default on Ultra" mean? We are running on
> UltraSPARC III but still triggered this problem. Is this a GCC bulid option or
> do we need the correct compile options for our software to make sure the
> correct specs are used? For now we patch the Makefile to use -fno-peephole.

It means that, if --with-cpu=ultrasparc or --with-cpu=ultrasparc3 is specified
at configure time, or -mcpu=ultrasparc or -mcpu=ultrasparc3 is specified at
compile time, -fno-peephole is effectively enabled for the SPARC back-end.

Note that --with-cpu=ultrasparc3 and -mcpu=ultrasparc3 are stupidly broken on
Solaris (the code doesn't assemble) in GCC 3.3.x and earlier, but not on Linux.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113

[Bug target/14113] [3.3/3.4 regression] -O2 corrupts stack

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-02-12 13:38 -------
Subject: Bug 14113

CVSROOT:	/cvs/gcc
Module name:	gcc
Branch: 	gcc-3_4-branch
Changes by:	ebotcazou@gcc.gnu.org	2004-02-12 13:38:16

Modified files:
	gcc            : ChangeLog 
	gcc/config/sparc: sparc.md 

Log message:
	PR target/14113
	Backport from mainline:
	
	2004-02-03  Eric Botcazou  <ebotcazou@libertysurf.fr>
	
	* config/sparc/sparc.md (call followed by jump define_peephole's):
	Delete.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&only_with_tag=gcc-3_4-branch&r1=2.2326.2.182&r2=2.2326.2.183
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/config/sparc/sparc.md.diff?cvsroot=gcc&only_with_tag=gcc-3_4-branch&r1=1.196.2.1&r2=1.196.2.2



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113

[Bug target/14113] [3.3/3.4 regression] -O2 corrupts stack

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-02-12 13:42 -------
Subject: Bug 14113

CVSROOT:	/cvs/gcc
Module name:	gcc
Branch: 	gcc-3_3-branch
Changes by:	ebotcazou@gcc.gnu.org	2004-02-12 13:42:51

Modified files:
	gcc            : ChangeLog 
	gcc/config/sparc: sparc.md 

Log message:
	PR target/14113
	Backport from mainline:
	
	2004-02-03  Eric Botcazou  <ebotcazou@libertysurf.fr>
	
	* config/sparc/sparc.md (call followed by jump define_peephole's):
	Delete.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&only_with_tag=gcc-3_3-branch&r1=1.16114.2.900&r2=1.16114.2.901
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/config/sparc/sparc.md.diff?cvsroot=gcc&only_with_tag=gcc-3_3-branch&r1=1.181.4.9&r2=1.181.4.10



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113

[Bug target/14113] [3.3/3.4 regression] -O2 corrupts stack

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2004-02-12 13:45 -------
See http://gcc.gnu.org/ml/gcc-patches/2004-02/msg01094.html


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|3.3.4                       |3.3.3


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14113