[Bug fortran/96312] New: Reallocation on assignment uses undefined variables

[Bug fortran/96312] New: Reallocation on assignment uses undefined variables

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96312

            Bug ID: 96312
           Summary: Reallocation on assignment uses undefined variables
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: fortran
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tkoenig at gcc dot gnu.org
  Target Milestone: ---

See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77504#c19 .

[Bug fortran/96312] Reallocation on assignment uses undefined variables

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96312

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=77504
                 CC|                            |dev-zero at gentoo dot org

--- Comment #1 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
Slightly sanitized test case from the originating PR, by Tiziano Müller:

$ cat moda.f90
module moda
contains
   PURE SUBROUTINE funca(arr, sz)
      REAL, ALLOCATABLE, DIMENSION(:, :), INTENT(OUT) :: arr
      integer, intent(in) :: sz
      allocate(arr(sz, sz))
      arr(:, :) = 0.
   END SUBROUTINE
end module
$ cat modb.f90
module modc
    use moda, only: funca
contains
   PURE SUBROUTINE funcb(oarr)
      REAL, DIMENSION(:), INTENT(OUT)    :: oarr
      REAL, ALLOCATABLE, DIMENSION(:, :) :: arr
      real, allocatable, dimension(:) :: tmp
      CALL funca(arr, ubound(oarr, 1))
      tmp = matmul(transpose(arr),oarr)
      oarr = tmp*1.
   END SUBROUTINE funcb
end module

Compiled with -O1 -Wall, this yields
$ gfortran -Wall -O1 -c moda.f90 modb.f90
modb.f90:8:0:

    8 |       oarr = MATMUL(TRANSPOSE(arr), oarr)*1.
      | 
Warning: '__var_1_matmul.dim[0].lbound' is used uninitialized in this function
[-Wuninitialized]
modb.f90:8:0: Warning: '__var_1_matmul.dim[0].ubound' is used uninitialized in
this function [-Wuninitialized]


Front-end optimization introduces an extra assignment statement here,
this version also shows the warning:

$ cat modc.f90
module modc
    use moda, only: funca
contains
   PURE SUBROUTINE funcb(oarr)
      REAL, DIMENSION(:), INTENT(OUT)    :: oarr
      REAL, ALLOCATABLE, DIMENSION(:, :) :: arr
      real, allocatable, dimension(:) :: tmp
      CALL funca(arr, ubound(oarr, 1))
      tmp = matmul(transpose(arr), oarr)
      oarr = tmp*1.
   END SUBROUTINE funcb
end module

The warning seems to be correct:

$ grep __var_1_matmul modb.f90.004t.original 
        struct array01_real(kind=4) __var_1_matmul;
            __var_1_matmul.data = 0B;
              __var_1_matmul.dtype = {.elem_len=4, .rank=1, .type=3};
              D.3969 = __var_1_matmul;
              D.3970 = (real(kind=4)[0:] * restrict) __var_1_matmul.data == 0B;
              __builtin_free ((void *) __var_1_matmul.data);
              __var_1_matmul.data = D.3969.data;
              D.3971 = ((__var_1_matmul.dim[0].lbound - D.3969.dim[0].lbound) -
__var_1_matmul.dim[0].ubound) + D.3969.dim[0].ubound != 0;
              D.3973 = D.3972 ? 1 : __var_1_matmul.dim[0].lbound;
              __var_1_matmul.dim[0].lbound = D.3973;
              __var_1_matmul.dim[0].ubound = D.3969.dim[0].ubound + D.3973;
              __var_1_matmul.dim[0].stride = 1;
              __var_1_matmul.offset = D.3974;
              D.4003 = (real(kind=4)[0:] * restrict) __var_1_matmul.data;
              D.4004 = __var_1_matmul.offset;
              D.4005 = __var_1_matmul.dim[0].lbound;
              D.4006 = __var_1_matmul.dim[0].ubound;
            if ((real(kind=4)[0:] * restrict) __var_1_matmul.data != 0B)
                __builtin_free ((void *) __var_1_matmul.data);
                (real(kind=4)[0:] * restrict) __var_1_matmul.data = 0B;

So, something seems to be wrong with reallocation on assignment here.

[Bug fortran/96312] [10/11 Regression] Reallocation on assignment uses undefined variables

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96312

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|Reallocation on assignment  |[10/11 Regression]
                   |uses undefined variables    |Reallocation on assignment
                   |                            |uses undefined variables
   Target Milestone|---                         |9.4

--- Comment #2 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
$ ~/Gcc/9-bin/gcc/f951 -quiet -O -Wall modc.f90
$ ~/Gcc/10-bin/gcc/f951 -quiet -O -Wall modc.f90
modc.f90:9:0:

    9 |       tmp = matmul(transpose(arr), oarr)
      | 
Warning: 'tmp.dim[0].lbound' is used uninitialized in this function
[-Wuninitialized]
modc.f90:9:0: Warning: 'tmp.dim[0].ubound' is used uninitialized in this
function [-Wuninitialized]
$ ~/Gcc/trunk-bin/gcc/f951 -quiet -O -Wall modc.f90
modc.f90:9:0:

    9 |       tmp = matmul(transpose(arr), oarr)
      | 
Warning: 'tmp.dim[0].lbound' is used uninitialized [-Wuninitialized]
modc.f90:7:44:

    7 |       real, allocatable, dimension(:) :: tmp
      |                                            ^
note: 'tmp' declared here
modc.f90:9:0:

    9 |       tmp = matmul(transpose(arr), oarr)
      | 
Warning: 'tmp.dim[0].ubound' is used uninitialized [-Wuninitialized]
modc.f90:7:44:

    7 |       real, allocatable, dimension(:) :: tmp
      |                                            ^
note: 'tmp' declared here

[Bug fortran/96312] [10/11 Regression] Reallocation on assignment uses undefined variables

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96312

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|9.4                         |10.3

[Bug fortran/96312] [10/11 Regression] Reallocation on assignment uses undefined variables

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96312

--- Comment #3 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
Let's see what bisection brings.

[Bug fortran/96312] [10/11 Regression] Reallocation on assignment uses undefined variables

[dominiq at lps dot ens.fr]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96312

Dominique d'Humieres <dominiq at lps dot ens.fr> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
           Priority|P3                          |P4
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2020-07-24

[Bug fortran/96312] [10/11 Regression] Reallocation on assignment uses undefined variables

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96312

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenther at suse dot de

--- Comment #4 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
Bisection points towards

2092f134b7180cd2542cff93bd8a876b3e59a77b is the first bad commit
commit 2092f134b7180cd2542cff93bd8a876b3e59a77b
Author: Richard Biener <rguenther@suse.de>
Date:   Wed May 15 12:57:32 2019 +0000

    tree-into-ssa.c (pass_build_ssa::execute): Run update_address_taken before
going into SSA.

    2019-05-15  Richard Biener  <rguenther@suse.de>

        * tree-into-ssa.c (pass_build_ssa::execute): Run
        update_address_taken before going into SSA.

    From-SVN: r271209

I tested this (in the bisect script) for the occurrence of the
warning with

$HOME/trunk-bin/gcc/f951 -quiet moda.f90
$HOME/trunk-bin/gcc/f951 -quiet -Wall -O3  modc.f90 -Werror

so it is likely that this patch just started issuing a warning
for a pre-existing bug in the front end.

[Bug fortran/96312] [10/11 Regression] Reallocation on assignment uses undefined variables

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96312

--- Comment #5 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
(In reply to Thomas Koenig from comment #4)

> so it is likely that this patch just started issuing a warning
> for a pre-existing bug in the front end.

That is indeed the case.  Grepping for tmp in the modc.f90 dump yields

  struct array01_real(kind=4) tmp;
      tmp.data = 0B;
        tmp.dtype = {.elem_len=4, .rank=1, .type=3};
        D.2309 = tmp;
        D.2310 = (real(kind=4)[0:] * restrict) tmp.data == 0B;
        __builtin_free ((void *) tmp.data);
        tmp.data = D.2309.data;
        D.2311 = ((tmp.dim[0].lbound - D.2309.dim[0].lbound) -
tmp.dim[0].ubound) + D.2309.dim[0].ubound != 0;
        D.2313 = D.2312 ? 1 : tmp.dim[0].lbound;
        tmp.dim[0].lbound = D.2313;
        tmp.dim[0].ubound = D.2309.dim[0].ubound + D.2313;
        tmp.dim[0].stride = 1;
        tmp.offset = D.2314;
        D.2343 = (real(kind=4)[0:] * restrict) tmp.data;
        D.2344 = tmp.offset;
        D.2345 = tmp.dim[0].lbound;
        D.2346 = tmp.dim[0].ubound;
      if ((real(kind=4)[0:] * restrict) tmp.data != 0B)
          __builtin_free ((void *) tmp.data);
          (real(kind=4)[0:] * restrict) tmp.data = 0B;

so there is no way that tmp.dim[0].lbound would have gotten
a value - previous version just didn't see it.

[Bug fortran/96312] [10/11 Regression] Reallocation on assignment uses undefined variables

[pault at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96312

Paul Thomas <pault at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |pault at gcc dot gnu.org
                 CC|                            |pault at gcc dot gnu.org

--- Comment #6 from Paul Thomas <pault at gcc dot gnu.org> ---
Created attachment 49031
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49031&action=edit
Patch for the PR

This is quite straightforward. The patch makes sure that references to the
bounds only occur if the lhs is allocated on arrival at offending statement.

It even regtests OK :-)

Paul

[Bug fortran/96312] [10/11 Regression] Reallocation on assignment uses undefined variables

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96312

--- Comment #7 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Paul Thomas <pault@gcc.gnu.org>:

https://gcc.gnu.org/g:abb276d0eca218e62e5ed50babf12ff544250759

commit r11-2625-gabb276d0eca218e62e5ed50babf12ff544250759
Author: Paul Thomas <pault@gcc.gnu.org>
Date:   Mon Aug 10 06:22:22 2020 +0100

    This patch fixes PR96312. Cures a used uninitialized warning.

    2020-08-10  Paul Thomas  <pault@gcc.gnu.org>

    gcc/fortran
            PR fortran/96312
            * trans-expr.c (fcncall_realloc_result): Only compare shapes if
            lhs was allocated..

    gcc/testsuite/
            PR fortran/96312
            * gfortran.dg/pr96312.f90: New test.

[Bug fortran/96312] [10/11 Regression] Reallocation on assignment uses undefined variables

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96312

--- Comment #8 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-10 branch has been updated by Thomas Kथà¤nig
<tkoenig@gcc.gnu.org>:

https://gcc.gnu.org/g:9040b14da7070f30af6d2814097fb7ea1a91707d

commit r10-8603-g9040b14da7070f30af6d2814097fb7ea1a91707d
Author: Paul Thomas <pault@gcc.gnu.org>
Date:   Mon Aug 10 06:22:22 2020 +0100

    This patch fixes PR96312. Cures a used uninitialized warning.

    2020-08-10  Paul Thomas  <pault@gcc.gnu.org>

    gcc/fortran
            PR fortran/96312
            * trans-expr.c (fcncall_realloc_result): Only compare shapes if
            lhs was allocated..

    gcc/testsuite/
            PR fortran/96312
            * gfortran.dg/pr96312.f90: New test.

    (cherry picked from commit abb276d0eca218e62e5ed50babf12ff544250759)

[Bug fortran/96312] [10/11 Regression] Reallocation on assignment uses undefined variables

[pault at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96312

Paul Thomas <pault at gcc dot gnu.org> changed:

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

--- Comment #9 from Paul Thomas <pault at gcc dot gnu.org> ---
Thanks for the backport, Thomas.

Fixed on both branches.

Paul