[Bug fortran/48352] New: [4.7 Regression] segfault in fortran/frontend-passes.c

[Bug fortran/48352] New: [4.7 Regression] segfault in fortran/frontend-passes.c

[Joost.VandeVondele at pci dot uzh.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48352

           Summary: [4.7 Regression] segfault in fortran/frontend-passes.c
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: fortran
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: Joost.VandeVondele@pci.uzh.ch


The following testcase started failing a couple of days ago:

MODULE cp_dbcsr_types
  TYPE cp_dbcsr_p_type
  END TYPE cp_dbcsr_p_type
CONTAINS
SUBROUTINE ep_qs_set(ep_qs_env, dH_coeffs_ptr, dS_coeffs_ptr, error)
    TYPE(cp_dbcsr_p_type), DIMENSION(:), &
      OPTIONAL, POINTER                      :: dH_coeffs_ptr, dS_coeffs_ptr
              DO i=1,MIN(SIZE(dS_coeffs_ptr),SIZE(dS_coeffs_ptr))
              END DO
END SUBROUTINE ep_qs_set
END MODULE

with the following segfault. Needs gfortran -O1.

Program received signal SIGSEGV, Segmentation fault.
gfc_expr_walker (e=0x18, exprfn=0x5bd5b0 <cfe_expr_0>, data=0x0) at
../../gcc/gcc/fortran/frontend-passes.c:741
741       while (*e)
(gdb) bt
#0  gfc_expr_walker (e=0x18, exprfn=0x5bd5b0 <cfe_expr_0>, data=0x0) at
../../gcc/gcc/fortran/frontend-passes.c:741
#1  0x00000000005bdabc in gfc_code_walker (c=0x15082d0, codefn=0x5bc9a0
<cfe_code>, exprfn=0x5bd5b0 <cfe_expr_0>, data=0x0)
    at ../../gcc/gcc/fortran/frontend-passes.c:1039
#2  0x00000000005be73b in optimize_namespace (ns=0x1507ba0) at
../../gcc/gcc/fortran/frontend-passes.c:338
#3  0x00000000005be768 in optimize_namespace (ns=0x1507ba0) at
../../gcc/gcc/fortran/frontend-passes.c:342
#4  0x00000000005be7b3 in gfc_run_passes (ns=0x1503d40) at
../../gcc/gcc/fortran/frontend-passes.c:69
#5  0x000000000052aee8 in gfc_parse_file () at
../../gcc/gcc/fortran/parse.c:4368
#6  0x0000000000564616 in gfc_be_parse_file () at
../../gcc/gcc/fortran/f95-lang.c:250
#7  0x000000000086fa5c in compile_file (argc=15, argv=0x7fffffffdc88) at
../../gcc/gcc/toplev.c:579
#8  do_compile (argc=15, argv=0x7fffffffdc88) at ../../gcc/gcc/toplev.c:1900
#9  toplev_main (argc=15, argv=0x7fffffffdc88) at ../../gcc/gcc/toplev.c:1963
#10 0x00007ffff661cb7d in __libc_start_main () from /lib64/libc.so.6
#11 0x00000000004c4e49 in _start () at ../sysdeps/x86_64/elf/start.S:113

[Bug fortran/48352] [4.7 Regression] segfault in fortran/frontend-passes.c

[burnus at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48352

Tobias Burnus <burnus at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |ice-on-valid-code
                 CC|                            |burnus at gcc dot gnu.org,
                   |                            |tkoenig at gcc dot gnu.org
   Target Milestone|---                         |4.7.0

[Bug fortran/48352] [4.7 Regression] segfault in fortran/frontend-passes.c

[tkoenig at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48352

Thomas Koenig <tkoenig at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2011.03.30 20:43:03
     Ever Confirmed|0                           |1

--- Comment #1 from Thomas Koenig <tkoenig at gcc dot gnu.org> 2011-03-30 20:43:03 UTC ---
Really strange.

gdb shows

Program received signal SIGSEGV, Segmentation fault.
gfc_expr_walker (e=0x18, exprfn=0x597fc0 <cfe_expr_0>, data=0x0)
    at ../../trunk/gcc/fortran/frontend-passes.c:752
752       while (*e)
(gdb) up
#1  0x00000000005984ec in gfc_code_walker (c=0x1452b30, codefn=0x597330
<cfe_code>,
    exprfn=0x597fc0 <cfe_expr_0>, data=0x0) at
../../trunk/gcc/fortran/frontend-passes.c:1050
1050                      WALK_SUBEXPR ((*c)->ext.omp_clauses->chunk_size);
(gdb) p (*c)->op
$1 = EXEC_ASSIGN
(gdb) p (*c)->ext.omp_clauses
$2 = (gfc_omp_clauses *) 0x0

Why are we in the section dealing with OMP stuff when handling an EXEC_ASSIGN?
We should not get there at all.

[Bug fortran/48352] [4.7 Regression] segfault in fortran/frontend-passes.c

[jvdelisle at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48352

Jerry DeLisle <jvdelisle at gcc dot gnu.org> changed:

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

--- Comment #2 from Jerry DeLisle <jvdelisle at gcc dot gnu.org> 2011-03-31 04:56:02 UTC ---
This fixes it.

Index: gcc/fortran/frontend-passes.c
===================================================================
--- gcc/fortran/frontend-passes.c    (revision 171769)
+++ gcc/fortran/frontend-passes.c    (working copy)
@@ -334,6 +334,8 @@ cfe_code (gfc_code **c, int *walk_subtrees ATTRIBU
 static void
 optimize_namespace (gfc_namespace *ns)
 {
+  if (!ns->code)
+    return;

   current_ns = ns;

Regression tested.

trim_optimize_2.f90 fails but I do not know if this is because the patterns are
changed or this patch is just disabling the optimizations.

Also moving the check down a little bit still works, as follows.

Index: gcc/fortran/frontend-passes.c
===================================================================
--- gcc/fortran/frontend-passes.c    (revision 171769)
+++ gcc/fortran/frontend-passes.c    (working copy)
@@ -334,11 +334,14 @@ cfe_code (gfc_code **c, int *walk_subtrees ATTRIBU
 static void
 optimize_namespace (gfc_namespace *ns)
 {
-
   current_ns = ns;

   gfc_code_walker (&ns->code, cfe_code, cfe_expr_0, NULL);
+
   gfc_code_walker (&ns->code, optimize_code, optimize_expr, NULL);
+  
+  if (!ns->code)
+    return;

   for (ns = ns->contained; ns; ns = ns->sibling)
     optimize_namespace (ns);

I think somewhere while walking the code, the code pointer is getting bashed or
for this test case, it really is null.

Hope this at least gives some hints.

[Bug fortran/48352] [4.7 Regression] segfault in fortran/frontend-passes.c

[jvdelisle at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48352

--- Comment #3 from Jerry DeLisle <jvdelisle at gcc dot gnu.org> 2011-03-31 12:54:26 UTC ---
The regression is caused by r171207.

[Bug fortran/48352] [4.7 Regression] segfault in fortran/frontend-passes.c

[burnus at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48352

--- Comment #4 from Tobias Burnus <burnus at gcc dot gnu.org> 2011-03-31 13:05:29 UTC ---
(In reply to comment #3)
> The regression is caused by r171207.

That's http://gcc.gnu.org/ml/gcc-cvs/2011-03/msg00630.html

Author: tkoenig
Date: Mon Mar 21 07:14:42 2011
New Revision: 171207

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=171207
Log:
2010-03-21  Thomas Koenig

    PR fortran/22572
[...]

[Bug fortran/48352] [4.7 Regression] segfault in fortran/frontend-passes.c

[Joost.VandeVondele at pci dot uzh.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48352

--- Comment #5 from Joost VandeVondele <Joost.VandeVondele at pci dot uzh.ch> 2011-03-31 13:25:27 UTC ---
reduced:

INTEGER, DIMENSION(:), POINTER :: a
DO I=1,MIN(SIZE(a),SIZE(a))
ENDDO
END

[Bug fortran/48352] [4.7 Regression] segfault in fortran/frontend-passes.c

[jvdelisle at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48352

--- Comment #6 from Jerry DeLisle <jvdelisle at gcc dot gnu.org> 2011-03-31 15:01:31 UTC ---
I am a little suspicious of this:

 static int count_arglist;

+/* Pointer to an array of gfc_expr ** we operate on, plus its size
+   and counter.  */
+
+static gfc_expr ***expr_array;
+static int expr_size, expr_count;


It could be getting walked on or optimized away by the optimizers.  Is there
another way to do this that is less indirect or relies less on static.

Just seems too tricky.  I do not have time to go further with this.

[Bug fortran/48352] [4.7 Regression] segfault in fortran/frontend-passes.c

[tkoenig at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48352

Thomas Koenig <tkoenig at gcc dot gnu.org> changed:

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

--- Comment #7 from Thomas Koenig <tkoenig at gcc dot gnu.org> 2011-03-31 18:31:22 UTC ---
(In reply to comment #5)
> reduced:
> 
> INTEGER, DIMENSION(:), POINTER :: a
> DO I=1,MIN(SIZE(a),SIZE(a))
> ENDDO
> END

Thanks, Joost.

Slightly more reduced test case (from the point of
stepping through the code):

interface
   pure integer function f()
   end function f
end interface
DO I=1,min(f(),f())
ENDDO
END

Somehow, the insertion of the statement with the extra variable doesn't
work for DO loops.

[Bug fortran/48352] [4.7 Regression] segfault in fortran/frontend-passes.c

[Joost.VandeVondele at pci dot uzh.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48352

--- Comment #8 from Joost VandeVondele <Joost.VandeVondele at pci dot uzh.ch> 2011-04-01 06:17:51 UTC ---
BTW, from this experience, it would be great to have the frontend optimizations
being protected by a switch (-f(no-)frontend-optimizations or similar) which
can default to true at -O1 and higher. 

First, this would provide an easy workaround for this kind of bugs.

Second, this would make sense in combination with LTO, where in principle one
could compile in the first pass with -O0, and only at link time with -Ox (this
makes sense from a compile time point of view). However, for this to generate
good code, the first pass should be -O0 -ffrontend-optimizations, otherwise the
middle end would have no chance.

[Bug fortran/48352] [4.7 Regression] segfault in fortran/frontend-passes.c

[tkoenig at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48352

--- Comment #9 from Thomas Koenig <tkoenig at gcc dot gnu.org> 2011-04-01 19:31:27 UTC ---
Author: tkoenig
Date: Fri Apr  1 19:31:23 2011
New Revision: 171849

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=171849
Log:
2011-04-01  Thomas Koenig  <tkoenig@gcc.gnu.org>

    PR fortran/48352
    * frontend-passes (cfe_register_funcs):  Don't
    register functions if they appear as iterators in DO loops.

2011-04-01  Thomas Koenig  <tkoenig@gcc.gnu.org>

    PR fortran/48352
    * gfortran.dg/function_optimize_3.f90:  New test.


Added:
    trunk/gcc/testsuite/gfortran.dg/function_optimize_3.f90
Modified:
    trunk/gcc/fortran/ChangeLog
    trunk/gcc/fortran/frontend-passes.c
    trunk/gcc/testsuite/ChangeLog

[Bug fortran/48352] [4.7 Regression] segfault in fortran/frontend-passes.c

[tkoenig at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48352

Thomas Koenig <tkoenig at gcc dot gnu.org> changed:

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

--- Comment #10 from Thomas Koenig <tkoenig at gcc dot gnu.org> 2011-04-01 19:32:35 UTC ---
Fixed, closing.