[Bug fortran/48412] New: [4.7 Regression] CP2K miscompiled due to some Fortran frontend pass

[Bug fortran/48412] New: [4.7 Regression] CP2K miscompiled due to some Fortran frontend pass

[Joost.VandeVondele at pci dot uzh.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48412

           Summary: [4.7 Regression] CP2K miscompiled due to some Fortran
                    frontend pass
           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


after the latest fix, CP2K now compiles at -O1, but is miscompiled. Manually
disabling the frontend passes like: 

gfc_run_passes (gfc_namespace *ns)
{
  if (0 && optimize)
    {

fixes the issue (and also running at -O0). The compiler was working fine before
the function optimization patch that went in a couple of days ago, so I suspect
that's the cause.

I have no other testcase than compiling CP2K and running an input like
cp2k/tests/QS/H2O.inp.

[Bug fortran/48412] [4.7 Regression] CP2K miscompiled due to some Fortran frontend pass

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48412

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

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

--- Comment #1 from Steven Bosscher <steven at gcc dot gnu.org> 2011-04-02 10:25:12 UTC ---
Joost, perhaps you can narrow it down further by using a debug counter.

You'd have to add a debug counter to dbgcnt.def, say "frontend1" 
Instead of "if (0 ...)", you would add'd do "if (dbg_cnt (frontend1) ...)".

Then compile with -fdbg-cnt=frontend1:N where N is a number, run, and see if
the bug occurs. The trick is to find the greatest value for N where the bug
occurs. This can be done with a binary search, as explained in dbgcnt.def.

[Bug fortran/48412] [4.7 Regression] CP2K miscompiled due to some Fortran frontend pass

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48412

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.7.0

[Bug fortran/48412] [4.7 Regression] CP2K miscompiled due to some Fortran frontend pass

[tkoenig at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48412

--- Comment #2 from Thomas Koenig <tkoenig at gcc dot gnu.org> 2011-04-03 17:28:14 UTC ---
Hi Joost,

the following patch

Index: frontend-passes.c
===================================================================
--- frontend-passes.c   (Revision 171913)
+++ frontend-passes.c   (Arbeitskopie)
@@ -279,6 +279,20 @@
   return result;
 }

+/* Warn about function removal, uncontitionally for now.  */
+
+static void
+warn_function_removal(gfc_expr *e)
+{
+  if (e->expr_type != EXPR_FUNCTION)
+    return;
+  if (e->value.function.esym)
+    gfc_warning ("Removing call to function %s at %L",
+                e->value.function.esym->name, &(e->where));
+  else if (e->value.function.isym)
+    gfc_warning ("Removing call to function %s at %L",
+                e->value.function.isym->name, &(e->where));
+}
 /* Callback function for the code walker for doing common function
    elimination.  This builds up the list of functions in the expression
    and goes through them to detect duplicates, which it then replaces
@@ -311,6 +325,8 @@
            {
              if (newvar == NULL)
                newvar = create_var (*(expr_array[i]));
+
+             warn_function_removal (*(expr_array[j]));
              gfc_free (*(expr_array[j]));
              *(expr_array[j]) = gfc_copy_expr (newvar);
            }

could maybe help you debug this by showing which functions get removed.

[Bug fortran/48412] [4.7 Regression] CP2K miscompiled due to some Fortran frontend pass

[Joost.VandeVondele at pci dot uzh.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48412

--- Comment #3 from Joost VandeVondele <Joost.VandeVondele at pci dot uzh.ch> 2011-04-04 09:06:32 UTC ---
(In reply to comment #2)
> Hi Joost,
> 
> the following patch
> 

hmm triggers 276 times on CP2K, quite a few seems things that also the ME would
capture. At least one is a bug in our code :-)

I'll see if can figure out the affected file.

[Bug fortran/48412] [4.7 Regression] CP2K miscompiled due to some Fortran frontend pass

[Joost.VandeVondele at pci dot uzh.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48412

--- Comment #4 from Joost VandeVondele <Joost.VandeVondele at pci dot uzh.ch> 2011-04-04 10:50:57 UTC ---
reduced testcase .... aborts at -O1, goes fine at -O0. Thanks for implementing
the warning, without this, it would have been very difficult to find.

INTEGER FUNCTION S1(m,ma,lx)
INTEGER :: m,ma,lx

IF (((m < 0).AND.(MODULO(ABS(ma-lx),2) == 1)).OR.&
    ((m > 0).AND.(MODULO(ABS(ma-lx),2) == 0))) THEN
   S1=1
ELSE
   S1=0
ENDIF

END FUNCTION

INTEGER :: s1
IF (S1(1,2,1).NE.0) CALL ABORT()
END

[Bug fortran/48412] [4.7 Regression] CP2K miscompiled due to some Fortran frontend pass

[Joost.VandeVondele at pci dot uzh.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48412

--- Comment #5 from Joost VandeVondele <Joost.VandeVondele at pci dot uzh.ch> 2011-04-04 11:49:15 UTC ---
the dumped code:

  __var_2 = __var_1 %[fl] 2;
  __var_1 = ABS_EXPR <*ma - *lx>;
  if (*m < 0 && __var_2 == 1 || *m > 0 && __var_2 == 0)

shows clearly what is wrong, var_2 is using var_1 before it is used.

[Bug fortran/48412] [4.7 Regression] CP2K miscompiled due to some Fortran frontend pass

[tkoenig at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48412

--- Comment #6 from Thomas Koenig <tkoenig at gcc dot gnu.org> 2011-04-04 20:22:27 UTC ---
Author: tkoenig
Date: Mon Apr  4 20:22:21 2011
New Revision: 171952

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=171952
Log:
2011-04-04  Thomas Koenig  <tkoenig@gcc.gnu.org>

    PR fortran/48412
    * frontend-passes (cfe_expr_0):  Reverse the order of going
    through the loops.

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

    PR fortran/48412
    * function_optimize_4.f90:  New test.


Added:
    trunk/gcc/testsuite/gfortran.dg/function_optimize_4.f90
Modified:
    trunk/gcc/fortran/ChangeLog
    trunk/gcc/fortran/frontend-passes.c
    trunk/gcc/testsuite/ChangeLog

[Bug fortran/48412] [4.7 Regression] CP2K miscompiled due to some Fortran frontend pass

[tkoenig at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48412

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

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

--- Comment #7 from Thomas Koenig <tkoenig at gcc dot gnu.org> 2011-04-04 20:27:32 UTC ---
Fixed on trunk, closing.

Thanks for the bug report and the analysis!