Bug 40873 - -fwhole-file -fwhole-program: Wrong decls cause too much to be optimized away
Summary: -fwhole-file -fwhole-program: Wrong decls cause too much to be optimized away
Status: RESOLVED FIXED
Alias: None
Product: gcc
Classification: Unclassified
Component: fortran (show other bugs)
Version: 4.5.0
: P3 normal
Target Milestone: ---
Assignee: Tobias Burnus
URL:
Keywords: rejects-valid, wrong-code
Depends on:
Blocks: 42361 40011 45087
  Show dependency treegraph
 
Reported: 2009-07-27 13:57 UTC by Tobias Burnus
Modified: 2010-07-27 08:46 UTC (History)
3 users (show)

See Also:
Host:
Target:
Build:
Known to work:
Known to fail: 4.6.0
Last reconfirmed: 2010-07-24 18:15:21


Attachments
Fix for this PR and PR40011 #42 (497 bytes, patch)
2010-05-24 12:31 UTC, Paul Thomas
Details | Diff
New trans-decl.c patch - seems to work well (2.02 KB, patch)
2010-07-26 13:25 UTC, Tobias Burnus
Details | Diff

Note You need to log in before you can comment on or make changes to this bug.
Description Tobias Burnus 2009-07-27 13:57:35 UTC
The option -fwhole-programs marks all functions in that file as local to that file, which facilitates inlining and dead-code elimination.

That of cause assumes that all declarations are properly unified, which should be the case with -fwhole-file. However, the following Polyhedron testcases fail
  ac, aermod, doduc, gas_dyn, linpk, mdbx, rnflow and test_fpu
at link time as too many functions have been elimated:

$ gfortran -fwhole-program -O1 linpk.f90
/tmp/ccWhxrXN.o: In function `MAIN__':
linpk.f90:(.text+0x759): undefined reference to `daxpy_'
linpk.f90:(.text+0x7c7): undefined reference to `daxpy_'
Comment 1 Tobias Burnus 2009-07-27 14:47:41 UTC
(Note: -fwhole-program implies -fwhole-file; the -O<n> option is required.)

Test case. Run with "-O1 -fwhole-program".
Fails with: test.f90:(.text+0x1b): undefined reference to `three_'

      program prog
        call one()
        call two()
      end program prog
      subroutine one()
        call three()
      end subroutine one
      subroutine two()
        call three()
      end subroutine two
      subroutine three()
      end subroutine three
Comment 2 Richard Biener 2009-07-28 13:12:15 UTC
We have the following cgraph nodes related to daxpy:

daxpy/17(-1) @0x7ffff5fc8700
  called by: dgesl/3 (1.00 per call) dgesl/3 (1.00 per call)
  calls:
dgefa/7(7) @0x7ffff5f7b100 174 time, 45 benefit 138 size, 36 benefit 20 bytes stack usage reachable body finalized inlinable
  called by: linpk/9 (1.00 per call)
  calls: daxpy/4 (1.00 per call) dscal/5 (1.00 per call) idamax/6 (1.00 per call)
daxpy/4(4) @0x7ffff5f66400 125 time, 56 benefit 125 size, 47 benefit reachable body finalized inlinable
  called by: dgefa/7 (1.00 per call)
  calls:
dgesl/3(3) @0x7ffff5f45f00 219 time, 52 benefit 165 size, 43 benefit 24 bytes stack usage reachable body finalized inlinable
  called by: linpk/9 (1.00 per call)
  calls: ddot/2 (1.00 per call) ddot/2 (1.00 per call) daxpy/17 (1.00 per call) daxpy/17 (1.00 per call)

where daxpy/17 is the one without a body (not merged with daxpy/4), called
by dgesl.  The call in dgefa is inlined (as single remaining call to a
then reclaimable function).

Confirmed with pauls latest patch applied.
Comment 3 Richard Biener 2009-09-22 15:34:34 UTC
Reconfirmed with current trunk.  If I move three to the top of the file I
even get

/tmp/ccGHHCCU.o: In function `main':
t.f90:(.text+0x1e): undefined reference to `three_'
t.f90:(.text+0x28): undefined reference to `three_'
collect2: ld returned 1 exit status

this makes -fwhole-file less useful than necessary.  LTO manages to fix up
the decls (thus, with -flto it builds and links fine with -fwhole-program).
Comment 4 Richard Biener 2009-09-22 15:42:35 UTC
Similar testcase from PR40011

SUBROUTINE c()
 CALL a()
END SUBROUTINE c

SUBROUTINE a()
END SUBROUTINE a

MODULE M
CONTAINS
 SUBROUTINE b()
   CALL c()
 END SUBROUTINE
END MODULE

USE M
CALL b()
END

> gfortran -fwhole-file -O t.f90
/tmp/ccW7Uhc6.o: In function `__m_MOD_b':
t.f90:(.text+0xa): undefined reference to `c_'
collect2: ld returned 1 exit status
Comment 5 Dominique d'Humieres 2010-05-15 19:53:30 UTC
A few comments:

(1) adding -flto or -fwhopr solves the linking problem for the polyhedron tests and the reduced one in comment #1.

(2) the test in comment #4 is different as it shows up for -fwhole-file and is not solved with -flto or -fwhopr.

(3) I have been puzzled by the results in http://users.physik.fu-berlin.de/~tburnus/gcc-trunk/benchmark/ for fatigue.f90. It is due to -fwhole-program:

[macbook] lin/test% gfc -O3 -ffast-math -fwhole-file -flto fatigue.f90
[macbook] lin/test% time a.out
...
9.223u 0.004s 0:09.23 99.8%	0+0k 0+0io 0pf+0w
[macbook] lin/test% gfc -O3 -ffast-math -fwhole-program fatigue.f90
[macbook] lin/test% time a.out
...
6.482u 0.004s 0:06.49 99.8%	0+0k 0+0io 0pf+0w

It would be interesting to understand why and to keep this nice speed up when fixing this pr.
Comment 6 Richard Biener 2010-05-16 10:53:32 UTC
(In reply to comment #5)
> A few comments:
> 
> (1) adding -flto or -fwhopr solves the linking problem for the polyhedron tests
> and the reduced one in comment #1.
> 
> (2) the test in comment #4 is different as it shows up for -fwhole-file and is
> not solved with -flto or -fwhopr.
> 
> (3) I have been puzzled by the results in
> http://users.physik.fu-berlin.de/~tburnus/gcc-trunk/benchmark/ for fatigue.f90.
> It is due to -fwhole-program:
> 
> [macbook] lin/test% gfc -O3 -ffast-math -fwhole-file -flto fatigue.f90
> [macbook] lin/test% time a.out
> ...
> 9.223u 0.004s 0:09.23 99.8%     0+0k 0+0io 0pf+0w
> [macbook] lin/test% gfc -O3 -ffast-math -fwhole-program fatigue.f90
> [macbook] lin/test% time a.out
> ...
> 6.482u 0.004s 0:06.49 99.8%     0+0k 0+0io 0pf+0w
> 
> It would be interesting to understand why and to keep this nice speed up when
> fixing this pr.

-fwhole-program enables -fwhole-file.

Comment 7 Dominique d'Humieres 2010-05-16 11:00:50 UTC
> -fwhole-program enables -fwhole-file.

Yes, but -fwhole-file does not enable -fwhole-program. All the polyhedron tests pass with -fwhole-file (and say -O3 -ffast-math), but the test in comment #4 fails with -whole-file.

Comment 8 rguenther@suse.de 2010-05-16 11:04:09 UTC
Subject: Re:  -fwhole-file -fwhole-program: Wrong decls
 cause too much to be optimized away

On Sun, 16 May 2010, dominiq at lps dot ens dot fr wrote:

> ------- Comment #7 from dominiq at lps dot ens dot fr  2010-05-16 11:00 -------
> > -fwhole-program enables -fwhole-file.
> 
> Yes, but -fwhole-file does not enable -fwhole-program. All the polyhedron tests
> pass with -fwhole-file (and say -O3 -ffast-math), but the test in comment #4
> fails with -whole-file.

-fwhole-file cannot enable -fwhole-program.  -fwhole-program says
to the optimizers that they do see the whole program - all callers
to functions defined in the current TU have to be visible (and
have correct callgraph edges, thus -fwhole-file).

You cant' compare -fwhole-file numbers to -fwhole-program numbers.
-fwhole-file is a correctness option, w/o it the Frontend generates
an invalid representation for the middle-end.

Richard.
Comment 9 Dominique d'Humieres 2010-05-16 11:16:40 UTC
> You cant' compare -fwhole-file numbers to -fwhole-program numbers.
> -fwhole-file is a correctness option, w/o it the Frontend generates
> an invalid representation for the middle-end.

Well, from what I saw running the polyhedron tests, -fwhole-file is more than a correctness option. I think it exposes more optimization opportunities to the middle end, giving faster executable for ac, aermod, and doduc. Note that adding -flto gives also some speed up for these tests. Due to this pr one cannot test the effect of -fwhole-program on half the tests. However using it for fatigue gives a quite large speed up I do not see for the seven other tests.

Comment 10 rguenther@suse.de 2010-05-16 11:21:15 UTC
Subject: Re:  -fwhole-file -fwhole-program: Wrong decls
 cause too much to be optimized away

On Sun, 16 May 2010, dominiq at lps dot ens dot fr wrote:

> ------- Comment #9 from dominiq at lps dot ens dot fr  2010-05-16 11:16 -------
> > You cant' compare -fwhole-file numbers to -fwhole-program numbers.
> > -fwhole-file is a correctness option, w/o it the Frontend generates
> > an invalid representation for the middle-end.
> 
> Well, from what I saw running the polyhedron tests, -fwhole-file is more than a
> correctness option. I think it exposes more optimization opportunities to the
> middle end, giving faster executable for ac, aermod, and doduc. Note that
> adding -flto gives also some speed up for these tests. Due to this pr one
> cannot test the effect of -fwhole-program on half the tests. However using it
> for fatigue gives a quite large speed up I do not see for the seven other
> tests.

It enables more optimization opportunities because calls inside the
unit are visible as such.  Without -fwhole-file nearly all calls
look like calls to external functions and local functions appear
unused.

Richard.
Comment 11 Paul Thomas 2010-05-20 13:51:23 UTC
(In reply to comment #10)
Am I right in thinking that -fwhole-file could be enabled by default, if this PR were to be fixed?  (The appropriate changes in the testsuite would have to be mad too.)

Paul
Comment 12 Paul Thomas 2010-05-24 12:31:52 UTC
Created attachment 20734 [details]
Fix for this PR and PR40011 #42

This patch regtests OK apart from some peculiarities in proc_ptr_comp_9.f90 and proc_ptr_23.f90, which fail to link with -g.  The problems do not appear to be associated with this patch, however.

My belief is that with this patch and corrections of the "legacy style" testsuite cases, -fwhole-file could be finally made the default.

Paul
Comment 13 Tobias Burnus 2010-05-26 14:28:22 UTC
Is this now fixed by the following commit? Or is something else to be done (additional fix, backporting, ...)?

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=159852
Log:
2010-05-26  Paul Thomas  <pault@gcc.gnu.org>

        PR fortran/40011
        * resolve.c (resolve_global_procedure): Resolve the gsymbol's
        namespace before trying to reorder the gsymbols.
Comment 14 Dominique d'Humieres 2010-05-26 14:41:01 UTC
> Is this now fixed by the following commit? Or is something else to be done
> (additional fix, backporting, ...)?

At least ac.f90 (probably all the items of the list) fails to link with -O -fwhole-program at revision 159855.
Comment 15 Tobias Burnus 2010-05-26 14:45:53 UTC
(In reply to comment #13)
> Is this now fixed by the following commit?

Answer: It is not. Comment 1 now works with "-fwhole-program -O1", but comment 0 and comment 4 still fail. (Though, they work with "-fwhole-file -O1".)

(In reply to comment #12)
> My belief is that with this patch and corrections of the "legacy style"
> testsuite cases, -fwhole-file could be finally made the default.

As there are no -fwhole-file failures in this PR (though -fwhole-file bugs, revealed through -fwhole-program), I think this PR does not prevent enabling -fwhole-file by default.
Comment 16 Francois-Xavier Coudert 2010-06-09 22:10:16 UTC
Another one that fails:


subroutine func (x)
  use demo
  integer :: x
  x = 999
end subroutine func

subroutine foo
  interface
    subroutine func(x)
      integer :: x
   end subroutine func
  end interface

  integer :: x
  call func(x)
  if (x /= 999) call abort ()

end subroutine foo

program test
  call foo
end
Comment 17 Joost VandeVondele 2010-07-24 18:15:21 UTC
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40873#c1 still fails with current trunk
Comment 18 Tobias Burnus 2010-07-24 19:13:42 UTC
(In reply to comment #17)
> comment 1 still fails with current trunk

(In reply to comment #1)
>       subroutine two()
>         call three()
>       end subroutine two
>
>       subroutine three()
>       end subroutine three

The problem is that one first generates code for "two" and thus calls
  gfc_get_extern_function_decl
to generate the decl for "three" - there is no existing declaration. As next step, one works on "three" and calls
   gfc_create_function_decl
which creates another declaration.


The test case in comment 1 and the one in comment 16 (with the bogus "use demo" comment out) worked with an initial version of the following patch. However, it give an ICE for the test in comment 4:
22.f90:5:0: internal compiler error: in build_function_decl, at fortran/trans-decl.c:1599

The solution was to make the newly generated procedure as global - which it surely is - otherwise, it were not accessible.

Index: gcc/fortran/trans-decl.c
===================================================================
--- gcc/fortran/trans-decl.c    (Revision 162502)
+++ gcc/fortran/trans-decl.c    (Arbeitskopie)
@@ -1411,9 +1411,19 @@ gfc_get_extern_function_decl (gfc_symbol
        && !sym->attr.use_assoc
        && !sym->backend_decl
        && gsym && gsym->ns
-       && ((gsym->type == GSYM_SUBROUTINE) || (gsym->type == GSYM_FUNCTION))
-       && gsym->ns->proc_name->backend_decl)
+       && ((gsym->type == GSYM_SUBROUTINE) || (gsym->type == GSYM_FUNCTION)))
     {
+      if (!gsym->ns->proc_name->backend_decl)
+       {
+         tree save_fn_decl = current_function_decl;
+         /* By construction, the external function cannot be
+            a contained procedure.  */
+         current_function_decl = NULL_TREE;
+         gfc_create_function_decl (gsym->ns);
+         current_function_decl = save_fn_decl;
+         gcc_assert (gsym->ns->proc_name->backend_decl);
+       }
+
       /* If the namespace has entries, the proc_name is the
         entry master.  Find the entry and use its backend_decl.
         otherwise, use the proc_name backend_decl.  */
Comment 19 Tobias Burnus 2010-07-24 20:16:15 UTC
And of course the patch won't work out of the box:

$ gfortran -O2 -g  gfortran.dg/entry_array_specs_2.f && ./a.out 
gfortran.dg/entry_array_specs_2.f:16:0: internal compiler error: in output_die, at dwarf2out.c:11046

Ditto for gfortran.dg/pr25603.f, gfortran.dg/proc_decl_2.f90, gfortran.fortran-torture/execute/mystery_proc.f90, gfortran.fortran-torture/execute/procarg.f90, 
gfortran.dg/loc_1.f90, gfortran.dg/value_test.f90, gfortran.dg/value_tests_f03.f90.

Probably something is wrong with the "current_function_decl = NULL_TREE;" - hopefully, it is easily fixable and does not require a completely different approach.


But at least:

(In reply to comment #0)
> the following Polyhedron testcases fail
>   ac, aermod, doduc, gas_dyn, linpk, mdbx, rnflow and test_fpu

Using -march=native -ffast-math -funroll-loops -ftree-loop-linear -O3 -fwhole-program -fwhole-file,

*ALMOST ALL* the polyhedron tests succeed - except for gas_dyn and test_fpu (for which there are still undefined references).
[Using "-g" one also gets issues with polyhedron tests.]
Comment 20 Tobias Burnus 2010-07-24 22:05:30 UTC
(In reply to comment #19)
> $ gfortran -O2 -g  gfortran.dg/entry_array_specs_2.f && ./a.out 
> gfortran.dg/entry_array_specs_2.f:16:0: internal compiler error:
> in output_die, at dwarf2out.c:11046

I had a closer look at loc_1.f90. With this patch, the functions fn and foo are actually cloned and inlined - without this does not happen. Consequently, the ICE goes ways with -fno-inline. Thus, the patch might actually be correct and the bug could be at a completely different place. Additionally, the patch might give some performance boost :-)

But first, the cause for the ICE has to be found. Any idea where to start searching?
Comment 21 Dominique d'Humieres 2010-07-25 10:03:39 UTC
With the patch in comment #18, I see all the failures reported in comment #19, plus

FAIL: gfortran.dg/whole_file_9.f90  -O  (internal compiler error)
FAIL: gfortran.dg/g77/13037.f  -O3 -g  (internal compiler error)

With -fwhole-program -O3 -g, most of the polyhedron tests fail either with "internal compiler error: in output_die, at dwarf2out.c:11046" or with a segmentation fault.

In addition without any option, I see several failures for codes with recursive functions, such as (from pr27613):

program test
interface
  function bad_stuff(n)
    integer :: bad_stuff (2)
    integer :: n(2)
  end function bad_stuff
   recursive function rec_stuff(n) result (tmp)
    integer :: n(2), tmp(2)
  end function rec_stuff
end interface
   integer :: res(2)
  res = bad_stuff((/-19,-30/))
  print *,  res
  if (any (res .ne. (/25,25/))) call abort ()
  if (any (rec_stuff((/-19,-30/)) .ne. (/25,25/))) call abort ()

end program test

  recursive function bad_stuff(n)
    integer :: bad_stuff (2)
    integer :: n(2), tmp(2)
    bad_stuff = rec_stuff (n)
    if((maxval (n)>0).and.(maxval (n) < 2)) then
      bad_stuff = bad_stuff + bad_stuff (maxval (n)+1) 
    endif
   entry rec_stuff(n) result (tmp)
    tmp=1
    if(maxval (n) < 5) then
      tmp = tmp + rec_stuff (n+1)
    endif
  end function bad_stuff
Comment 22 Tobias Burnus 2010-07-25 10:14:30 UTC
The patch in comment 18 causes a segfault (in gfc_generate_function_code for "cfun->function_end_locus = input_location" [Invalid write of size 4]) for the test case in PR 40011 comment 0 (the one after "Your patch fixes some Segmentation faults (a couple)") - the test case works otherwise. No -fwhole-program is needed.
Comment 23 Tobias Burnus 2010-07-26 13:25:42 UTC
Created attachment 21315 [details]
New trans-decl.c patch - seems to work well

New patch. Found the problem with the help of Jakub (thanks!); not yet regtested but it works with the previously failing examples.
Comment 24 Dominique d'Humieres 2010-07-26 17:02:12 UTC
With the patch in comment #23, the polyhedron tests gas_dyn.f90 and test_fpu.f90 do not link and compiling the test in http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31867#c6 gives an ICE (Segmentation fault). Otherwise all the other polyhedron tests compile with -O3 -g -fwhole-program and I did not see any miscompilation, regtested without regression. Nice work! thanks.
Comment 25 Tobias Burnus 2010-07-26 17:03:42 UTC
(In reply to comment #23)
> Created an attachment (id=21315) [edit]
> New trans-decl.c patch - seems to work well

Dominique has found a failure (segfault) with PR 31867 comment 6.

If one generates in gfc_get_extern_function_decl the code for "lensum", one finds that its argument "words" has locally the correct type:
  (gdb) p sym->formal->sym->as->type
  $4 = AS_ASSUMED_SHAPE
but the gsym has the wrong type
  (gdb) p gsym->ns->proc_name->formal->sym->as->type
  $10 = AS_DEFERRED
Thus, one enters the code path for descriptor-free arrays and crashes as UBOUND is NULL.

In principle, this should get fixed in resolve_formal_arglist. One problem is that if one enters find_arglists sym->ns != gfc_current_ns it fails.

But the actual problems seems to be in resolve_global_procedure. One has:

(gdb) p sym->attr.if_source
$27 = IFSRC_IFBODY
(gdb) p sym->formal->sym->as->type
$28 = AS_ASSUMED_SHAPE

That is: The symbol in the interface block of the module is resolved. But the gsym is not:

(gdb) p gsym->ns->resolved
$29 = 0
(gdb) p gsym->ns->proc_name->formal->sym->as->type
$30 = AS_DEFERRED

The following patch fixes the program. (Side remark, one could do more argument checking, cf. PR 45086.)

Index: gcc/fortran/resolve.c
===================================================================
--- gcc/fortran/resolve.c       (revision 162538)
+++ gcc/fortran/resolve.c       (working copy)
@@ -1816,7 +1816,8 @@ resolve_global_procedure (gfc_symbol *sy
     gfc_global_used (gsym, where);

   if (gfc_option.flag_whole_file
-       && sym->attr.if_source == IFSRC_UNKNOWN
+       && (sym->attr.if_source == IFSRC_UNKNOWN
+           || sym->attr.if_source == IFSRC_IFBODY)
        && gsym->type != GSYM_UNKNOWN
        && gsym->ns
        && gsym->ns->resolved != -1
@@ -1902,7 +1903,7 @@ resolve_global_procedure (gfc_symbol *sy
                   sym->name, &sym->declared_at, gfc_typename (&sym->ts),
                   gfc_typename (&def_sym->ts));

-      if (def_sym->formal)
+      if (def_sym->formal && sym->attr.if_source != IFSRC_IFBODY)
        {
          gfc_formal_arglist *arg = def_sym->formal;
          for ( ; arg; arg = arg->next)
Comment 26 Dominique d'Humieres 2010-07-26 21:00:20 UTC
The patch in comment #25 fixes the ICE for PR 31867 comment 6, but causes also several regressions with "... must have an explicit interface" errors:


FAIL: gfortran.dg/allocatable_function_1.f90 (test for excess errors)
FAIL: gfortran.dg/allocatable_function_3.f90  (test for excess errors)
FAIL: gfortran.dg/char_result_3.f90  (test for excess errors)
FAIL: gfortran.dg/char_result_4.f90  (test for excess errors)
FAIL: gfortran.dg/f2c_6.f90  (test for excess errors)
FAIL: gfortran.dg/import6.f90  (test for excess errors)
FAIL: gfortran.dg/pointer_check_6.f90  (test for excess errors)
FAIL: gfortran.dg/value_tests_f03.f90  (test for excess errors)
FAIL: gfortran.fortran-torture/execute/entry_7.f90 compilation
Comment 27 Tobias Burnus 2010-07-27 08:44:39 UTC
Subject: Bug 40873

Author: burnus
Date: Tue Jul 27 08:44:22 2010
New Revision: 162557

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=162557
Log:
2010-07-26  Tobias Burnus  <burnus@net-b.de>

        PR fortran/40873
        * trans-decl.c (gfc_get_extern_function_decl): Fix generation
        for functions which are later in the same file.
        (gfc_create_function_decl, build_function_decl,
        build_entry_thunks): Add global argument.
        * trans.c (gfc_generate_module_code): Update
        gfc_create_function_decl call.
        * trans.h (gfc_create_function_decl): Update prototype.
        * resolve.c (resolve_global_procedure): Also resolve for
        IFSRC_IFBODY.

2010-07-26  Tobias Burnus  <burnus@net-b.de>

        PR fortran/40873
        * gfortran.dg/whole_file_22.f90: New test.
        * gfortran.dg/whole_file_23.f90: New test.


Added:
    trunk/gcc/testsuite/gfortran.dg/whole_file_22.f90
    trunk/gcc/testsuite/gfortran.dg/whole_file_23.f90
Modified:
    trunk/gcc/fortran/ChangeLog
    trunk/gcc/fortran/resolve.c
    trunk/gcc/fortran/trans-decl.c
    trunk/gcc/fortran/trans.c
    trunk/gcc/fortran/trans.h
    trunk/gcc/testsuite/ChangeLog

Comment 28 Tobias Burnus 2010-07-27 08:46:19 UTC
FIXED on the trunk (4.6). Thanks for the reports, comments, patches, suggestions, and reviews!

See PR 45077, PR 45087, and PR 44945 for remaining -fwhole-(file,program) bugs.