[Patch, fortran] PR61459 - [4.8/4.9/4.10 Regression] segfault when assigning to allocatable function result from matmul result

[Patch, fortran] PR61459 - [4.8/4.9/4.10 Regression] segfault when assigning to allocatable function result from matmul result

[Paul Richard Thomas]

Dear All,

This PR is caused by a rather unfortunate cast for the pointer to the
return value from matmul:

      struct array1_real(kind=4) D.2357;
      D.2357 = *(struct array1_real(kind=4) *) __result;
      D.2357.data = 0B;
....snip....
      _gfortran_matmul_r4 (*(struct array1_real(kind=4) * *) &D.2357,
D.2384, D.2391, 0, 0, 0B);

Unsurprisingly, matmul_r4 segfaults because it is passed a NULL for
the descriptor.

The fix is obvious (ie use its natural kind) and results in:
      _gfortran_matmul_r4 (&D.2357, D.2384, D.2391, 0, 0, 0B);

Dominique points out that the patch fixes PR58883 as well.  I'll
enhance the testcase appropriately.

Unless anybody has any objections, I'll commit to trunk tonight and to
4.8 and 4.9 next weekend.

Cheers

Paul

2014-07-06  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/61459
* trans-expr.c (fcncall_realloc_result): Use the natural type
for the address expression of 'res_desc'.
2014-07-06  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/61459
* gfortran.dg/allocatable_function_8.f90 : New test

Index: gcc/fortran/trans-expr.c
===================================================================
*** gcc/fortran/trans-expr.c (revision 211867)
--- gcc/fortran/trans-expr.c (working copy)
*************** fcncall_realloc_result (gfc_se *se, int
*** 7299,7305 ****

    res_desc = gfc_evaluate_now (desc, &se->pre);
    gfc_conv_descriptor_data_set (&se->pre, res_desc, null_pointer_node);
!   se->expr = gfc_build_addr_expr (TREE_TYPE (se->expr), res_desc);

    /* Free the lhs after the function call and copy the result data to
       the lhs descriptor.  */
--- 7299,7305 ----

    res_desc = gfc_evaluate_now (desc, &se->pre);
    gfc_conv_descriptor_data_set (&se->pre, res_desc, null_pointer_node);
!   se->expr = gfc_build_addr_expr (NULL_TREE, res_desc);

    /* Free the lhs after the function call and copy the result data to
       the lhs descriptor.  */
Index: gcc/testsuite/gfortran.dg/allocatable_function_8.f90
===================================================================
*** gcc/testsuite/gfortran.dg/allocatable_function_8.f90 (revision 0)
--- gcc/testsuite/gfortran.dg/allocatable_function_8.f90 (working copy)
***************
*** 0 ****
--- 1,47 ----
+ ! { dg-do run }
+ ! Test the fix for PR61459.
+ !
+ ! Contributed by John Wingate  <johnww@tds.net>
+ !
+ module a
+
+    implicit none
+    private
+    public :: f_segfault, f_segfault_plus, f_workaround
+    integer, dimension(2,2) :: b = reshape([1,-1,1,1],[2,2])
+
+ contains
+
+    function f_segfault(x)
+       real, dimension(:), allocatable :: f_segfault
+       real, dimension(:), intent(in)  :: x
+       allocate(f_segfault(2))
+       f_segfault = matmul(b,x)
+    end function f_segfault
+
+ ! Sefaulted without the ALLOCATE as well.
+    function f_segfault_plus(x)
+       real, dimension(:), allocatable :: f_segfault_plus
+       real, dimension(:), intent(in)  :: x
+       f_segfault_plus = matmul(b,x)
+    end function f_segfault_plus
+
+    function f_workaround(x)
+       real, dimension(:), allocatable :: f_workaround
+       real, dimension(:), intent(in)  :: x
+       real, dimension(:), allocatable :: tmp
+       allocate(f_workaround(2),tmp(2))
+       tmp = matmul(b,x)
+       f_workaround = tmp
+    end function f_workaround
+
+ end module a
+
+ program main
+    use a
+    implicit none
+    real, dimension(2) :: x = 1.0, y
+    y = f_workaround (x)
+    if (any (f_segfault (x) .ne. y)) call abort
+    if (any (f_segfault_plus (x) .ne. y)) call abort
+ end program main

Re: [Patch, fortran] PR61459 - [4.8/4.9/4.10 Regression] segfault when assigning to allocatable function result from matmul result

[Paul Richard Thomas]

Dear Tobias,

The patch is so safe and, as you say, has no side effects that I will
commit to 4.9 first.

Cheers

Paul

On 7 July 2014 10:05, Tobias Burnus <tobias.burnus@physik.fu-berlin.de> wrote:
> Paul Richard Thomas wrote:
>> This PR is caused by a rather unfortunate cast for the pointer to the
>> return value from matmul:
> [...]
>> The fix is obvious (ie use its natural kind)
> [...]
>> Dominique points out that the patch fixes PR58883 as well.  I'll
>> enhance the testcase appropriately.
>
> Nice!
>
>> Unless anybody has any objections, I'll commit to trunk tonight and to
>> 4.8 and 4.9 next weekend.
>
> Looks good to me. As the patch shouldn't have side effects and as Jakub
> intents to do a 4.9.1 RC early this week, I'd prefer it the 4.9 patch
> would be committed earlier.
>
> Tobias



-- 
The knack of flying is learning how to throw yourself at the ground and miss.
       --Hitchhikers Guide to the Galaxy

Re: [Patch, fortran] PR61459 - [4.8/4.9/4.10 Regression] segfault when assigning to allocatable function result from matmul result

[Tobias Burnus]

Paul Richard Thomas wrote:
> This PR is caused by a rather unfortunate cast for the pointer to the
> return value from matmul:
[...]
> The fix is obvious (ie use its natural kind)
[...]
> Dominique points out that the patch fixes PR58883 as well.  I'll
> enhance the testcase appropriately.

Nice!

> Unless anybody has any objections, I'll commit to trunk tonight and to
> 4.8 and 4.9 next weekend.

Looks good to me. As the patch shouldn't have side effects and as Jakub
intents to do a 4.9.1 RC early this week, I'd prefer it the 4.9 patch
would be committed earlier.

Tobias