Created attachment 12841 [details]
main.f90 + 3 modules included my main

Confirmed. Definitely excessive.

if you try the example, f951 may exit with a segfault.  reason is this code takes a lot of stack space.

for tcsh, "ulimit stacksize unlimited"  was required.

(just to save 10 minutes for the next person who takes a look at this !!)

from what i see so far, the problem is in the .mod files, not in the reading of them.  there are hundreds of thousands of 'commons' defined, which is silly. 

any way to get source to the attached mod files ?


--bud

> from what i see so far, the problem is in the .mod files, not in the reading of
> them.  there are hundreds of thousands of 'commons' defined, which is silly. 

This gave me the idea to the code below.  Try compiling:


module mod0
  implicit none
  INCLUDE 'mpif.h'
  private
  public :: sub0
contains
  subroutine sub0 ()
  end subroutine sub0
end module mod0

module mod1
  use mod0, only : sub0
  implicit none
  private
  public :: sub1
contains
  subroutine sub1 ()
  end subroutine sub1
end module mod1

module mod2
  use mod0, only : sub0
  use mod1, only : sub1
  implicit none
  private
  public :: sub2
contains
  subroutine sub2 ()
  end subroutine sub2
end module mod2


with mpif.h cut down to:

      INTEGER MPI_BOTTOM
      COMMON /MPIPRIV/ MPI_BOTTOM

and look at the module files.  If you repeat the above,
the module files will explode.

Created attachment 12998 [details]
Module file size explosion demo

I have rewritten the demo code.  See how the size of
the module files explodes with gfortran.

perfect.  that seems to duplicate the problem.

the below patch looks like it fixes the problem.  any chance this could be tryed on the reported problem ?

--bud


Index: gcc/gcc/fortran/module.c
===================================================================
--- gcc/gcc/fortran/module.c    (revision 122310)
+++ gcc/gcc/fortran/module.c    (working copy)
@@ -3598,6 +3598,7 @@
 write_common (gfc_symtree *st)
 {
   gfc_common_head *p;
+  pointer_info *pinfo;
   const char * name;
   int flags;
 
@@ -3607,6 +3608,14 @@
   write_common (st->left);
   write_common (st->right);
 
+  /* only need to write a given symtree entry once for 
+     common blocks */ 
+  p = st->n.common; 
+  pinfo = find_pointer(p);
+
+  if (pinfo != NULL && pinfo->u.wsym.state != UNREFERENCED)
+    return;
+
   mio_lparen ();
 
   /* Write the unmangled name.  */
@@ -3614,7 +3623,6 @@
 
   mio_pool_string (&name);
 
-  p = st->n.common;
   mio_symbol_ref (&p->head);
   flags = p->saved ? 1 : 0;
   if (p->threadprivate) flags |= 2;

(In reply to comment #8)
> the below patch looks like it fixes the problem.  any chance this could be
> tryed on the reported problem ?

If somebody with sufficient resources can provide a binary
(like FX's snapshots), I will try to recompile the project.

-ha

Bud, I tried your patch, but it doesn't seem to make a difference.  Is it the right patch?

i think not.  i must have confued myself (rather easy to do!).  will dig through the source this weekend.

i can confirm the attached patch is wrong.

I've adjusted the title slightly to refer to the use of
COMMON blocks (in MPICH, OpenMPI) because the problem of
.mod size excursion depends only on the "nesting depth",
not on the actual size of the modules.  I hope this bug
will be fixed someday when somebody has a look either at
the module writing routines or at the way the structures
containing module information are held internally.

I am unable to compile a reasonably large project (just in
terms of the number of files, not total source code size!)
using MPI.
The problem appears even worse with OpenMPI than with MPICH.

I think the problem starts with:

module mod0
  integer mpi_bottom
  common /mpipriv/ mpi_bottom
end module mod0

module mod1
  use mod0
end module mod1

module mod2
  use mod0
  use mod1
end module mod2


Because in that case, mod2.mod already has two copies of the common:

  (('mpipriv' 2 0 0 'mpipriv') ('mpipriv' 2 0 0 'mpipriv'))

and I don't think that's desirable. I think that the module loading is actually wrong here: the code in gfc_get_common (match.c) takes special care to duplicate this common name by creating a unique name for it. While I believe that mangling is necessary, the mangled name shouldn't be unique but simply prefixed, so that of the same name are merged, while prevented to clash with the namespace of the use'ing procedure.

I'm regtesting the following patch:

Index: match.c
===================================================================
--- match.c     (revision 129869)
+++ match.c     (working copy)
@@ -2608,23 +2608,19 @@ gfc_common_head *
 gfc_get_common (const char *name, int from_module)
 {
   gfc_symtree *st;
-  static int serial = 0;
-  char mangled_name[GFC_MAX_SYMBOL_LEN + 1];
+  char mangled_name[GFC_MAX_SYMBOL_LEN + 12];

   if (from_module)
-    {
-      /* A use associated common block is only needed to correctly layout
-        the variables it contains.  */
-      snprintf (mangled_name, GFC_MAX_SYMBOL_LEN, "_%d_%s", serial++, name);
-      st = gfc_new_symtree (&gfc_current_ns->common_root, mangled_name);
-    }
-  else
-    {
-      st = gfc_find_symtree (gfc_current_ns->common_root, name);
-
-      if (st == NULL)
-       st = gfc_new_symtree (&gfc_current_ns->common_root, name);
-    }
+    /* A use associated common block is only needed to correctly layout
+       the variables it contains.  */
+    snprintf (mangled_name, GFC_MAX_SYMBOL_LEN, "_frommodule_%s", name);
+
+  st = gfc_find_symtree (gfc_current_ns->common_root,
+                        from_module ? mangled_name : name);
+
+  if (st == NULL)
+    st = gfc_new_symtree (&gfc_current_ns->common_root,
+                         from_module ? mangled_name : name);

   if (st->n.common == NULL)
     {

Subject: Re:  gfortran excessive memory usage with COMMON
 blocks in modules

fxcoudert at gcc dot gnu dot org wrote:
> Because in that case, mod2.mod already has two copies of the common:
> 
>   (('mpipriv' 2 0 0 'mpipriv') ('mpipriv' 2 0 0 'mpipriv'))
> 
> and I don't think that's desirable. I think that the module loading is actually
> wrong here: the code in gfc_get_common (match.c) takes special care to
> duplicate this common name by creating a unique name for it. While I believe
> that mangling is necessary, the mangled name shouldn't be unique but simply
> prefixed, so that of the same name are merged, while prevented to clash with
> the namespace of the use'ing procedure.

I wrote this code originally, and I agree with your analysis.

> Index: match.c
> ===================================================================
> --- match.c     (revision 129869)
> +++ match.c     (working copy)
> @@ -2608,23 +2608,19 @@ gfc_common_head *
>  gfc_get_common (const char *name, int from_module)
>  {
>    gfc_symtree *st;
> -  static int serial = 0;
> -  char mangled_name[GFC_MAX_SYMBOL_LEN + 1];
> +  char mangled_name[GFC_MAX_SYMBOL_LEN + 12];

Should be + 13 (need one char for '\0').

> +    /* A use associated common block is only needed to correctly layout
> +       the variables it contains.  */
> +    snprintf (mangled_name, GFC_MAX_SYMBOL_LEN, "_frommodule_%s", name);

GFC_MAX_SYMBOL_LEN + 12, otherwise you could create ambiguities with 
really long common names.  Previously this wasn't possible due to the 
serial number.

(In reply to comment #15)
> I wrote this code originally, and I agree with your analysis.

But regtesting doesn't agree with my analysis... in case of common with bind(c,name="..."), this patch hampers the diagnosis of commons with the same name but different labels. I've tried hard to get it rolling that way, because I agree it's cleaner, but I couldn't. I'll propose a different approach (a hack, to avoid writing twice the same combination of name and binding label) to the list when it finishes regtesting.

Subject: Re:  gfortran excessive memory usage with COMMON
 blocks in modules

fxcoudert at gcc dot gnu dot org wrote:
> ------- Comment #16 from fxcoudert at gcc dot gnu dot org  2007-11-09 23:59 -------
> (In reply to comment #15)
>> I wrote this code originally, and I agree with your analysis.
> 
> But regtesting doesn't agree with my analysis... in case of common with
> bind(c,name="..."), this patch hampers the diagnosis of commons with the same
> name but different labels. I've tried hard to get it rolling that way, because
> I agree it's cleaner, but I couldn't. I'll propose a different approach (a
> hack, to avoid writing twice the same combination of name and binding label) to
> the list when it finishes regtesting.

There was no BIND(C) when I wrote it :-)

I'll give it a look tomorrow.

Subject: Bug number PR 30285

A patch for this bug has been added to the patch tracker.
The mailing list url for the patch is http://gcc.gnu.org/ml/gcc-patches/2007-11/msg00552.html

Harald, can you test the patch at
  http://gcc.gnu.org/ml/gcc-patches/2007-11/msg00552.html
with your real-world program? Due to .mod changes, the "gfc-excessive.tar.gz" does not work. And for the other tests I have, the compile time does not significantly change (too much noise to see the effect).

Tobias,

> Harald, can you test the patch at
>   http://gcc.gnu.org/ml/gcc-patches/2007-11/msg00552.html
> with your real-world program?

I would love to test it with a i586/i686-compatible build.  (f951 should
be sufficient that can be used as a replacement in one of those regular
builds provided on FX's web page.  Unfortunately I don't have the
resourced to build gfortran myself.)

> And for the other tests I have, the compile time does not
> significantly change (too much noise to see the effect).

The essential problem was actually the virtual memory used by f951
when I last tried to compile the app with MPI enabled.  I think the
first "benchmark" is to have the module files of the reworked
example to all have the same size and essentially same contents.
From comment #14 I presume that this will indeed be the case.

Thanks,
-ha

(In reply to comment #20)
> I would love to test it with a i586/i686-compatible build.  (f951 should
> be sufficient that can be used as a replacement in one of those regular
> builds provided on FX's web page.  Unfortunately I don't have the
> resourced to build gfortran myself.)

I made a build of the patched compiler that you can download from http://www.coudert.name/tmp/gfortran-i686-linux-20071110.tar.gz

> I think the
> first "benchmark" is to have the module files of the reworked
> example to all have the same size and essentially same contents.
> From comment #14 I presume that this will indeed be the case.

Yes, it is the case. Please let us know how your real-life testing goes.

FX,
 
> I made a build of the patched compiler that you can download from
> http://www.coudert.name/tmp/gfortran-i686-linux-20071110.tar.gz

I successfully recompiled the app with OpenMPI enabled, which uses
several COMMON blocks.  Not only are the module files about a factor
100 smaller for me ;-), the compilation actually completes and the
executable works.  I am currently investigating the output, but the
problem described in this PR is solved for me.

Many thanks,
-ha

Subject: Bug number PR 30285

A patch for this bug has been added to the patch tracker.
The mailing list url for the patch is http://gcc.gnu.org/ml/gcc-patches/2007-11/msg00552.html

Subject: Bug 30285

Author: fxcoudert
Date: Sat Nov 17 13:46:53 2007
New Revision: 130257

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=130257
Log:
	PR fortran/30285
	* module.c (struct written_common, written_commons): New structure.
	(compare_written_commons, free_written_common, write_common_0):
	New functions.
	(write_common): Call recursive function write_common_0.

Modified:
    trunk/gcc/fortran/ChangeLog
    trunk/gcc/fortran/module.c

Fixed.