[Bug fortran/102619] [11/12/13/14/15 Regression] ICE in gfc_conv_descriptor_dtype, at fortran/trans-array.c:215 since r9-6493-g0e3088806577e805

[mikael at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102619

Mikael Morin  changed:

   What|Removed |Added

 CC||mikael at gcc dot gnu.org

--- Comment #8 from Mikael Morin  ---
The reason why the variable is not represented with an array descriptor is the
interface mapping mechanism to evaluate dummy array bounds.
gfc_add_interface_mapping replaces the normal variable decl with a
descriptorless (GFC_ARRAY_TYPE_P == 1) array.  The rank for those is
GFC_TYPE_ARRAY_RANK, which is #defined to (TYPE_LANG_SPECIFIC(node)->rank).
And lang_type declares the rank field with:
  int rank, corank;
so compile-time only rank.
gfc_set_interface_mapping_bounds uses a loop bounded by GFC_TYPE_ARRAY_RANK, so
it does nothing if the rank is -1 as with assumed rank.

So we expect a regular array descriptor from which we can get the runtime rank,
and we get a descriptorless array of rank -1 with unset bounds.
No wonder it doesn't work.

[Bug fortran/102619] [11/12/13/14/15 Regression] ICE in gfc_conv_descriptor_dtype, at fortran/trans-array.c:215 since r9-6493-g0e3088806577e805

[jvdelisle2 at gmail dot com via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102619

Jerry DeLisle  changed:

   What|Removed |Added

 CC||jvdelisle2 at gmail dot com

--- Comment #7 from Jerry DeLisle  ---
Interestingly, real :: x(:,:) works vs real :: x(..) in the function
declaration with your partial patch.

[Bug fortran/102619] [11/12/13/14/15 Regression] ICE in gfc_conv_descriptor_dtype, at fortran/trans-array.c:215 since r9-6493-g0e3088806577e805

[anlauf at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102619

--- Comment #6 from anlauf at gcc dot gnu.org ---
Created attachment 58302
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58302=edit
Partial patch

This change prevents the ICE and leads to a correct shape of the function
result for the following testcase:

program p
  implicit none
  real :: w(2,3)
  real, allocatable :: y(:)
  y = h(w)
  print *, size (y) ! Should print 6
  print *, y
  deallocate (y)
contains
  function h(x) result (g)
real :: x(..)
real :: g(product(shape(x)))
integer :: i
print *, shape (x)
print *, size (g)
g = [(real(i),i=1,size(g))]
print *, g
  end
end

After the patch this prints:

   2   3
   6
   1.   2.   3.   4.  
5.   6.
   1
   1.

So the function result is not allocated correctly.

[Bug fortran/102619] [11/12/13/14/15 Regression] ICE in gfc_conv_descriptor_dtype, at fortran/trans-array.c:215 since r9-6493-g0e3088806577e805

[anlauf at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102619

anlauf at gcc dot gnu.org changed:

   What|Removed |Added

 CC||anlauf at gcc dot gnu.org

--- Comment #5 from anlauf at gcc dot gnu.org ---
The ICE occurs for both SUM and PRODUCT of SHAPE,LBOUND,UBOUND of the
assumed-rank dummy.

The ICE points to gfc_conv_intrinsic_arith:

4929  gfc_conv_ss_startstride ();

The difference between assumed-shape vs. assumed-rank argument is:

(gdb) p *loop->ss->info->expr->shape->_mp_d
$86 = 1

vs.

(gdb) p loop->ss->info->expr->shape
$88 = (mpz_t *) 0x0

Now gfc_conv_ss_startstride should be able to handle assumed-rank,
but then we die here:

(gdb) l 4811,4826
4811arg = expr->value.function.actual->expr;
4812if (arg->rank == -1)
4813  {
4814gfc_se se;
4815tree rank, tmp;
4816
4817/* The rank (hence the return value's shape) is
unknown,
4818   we have to retrieve it.  */
4819gfc_init_se (, NULL);
4820se.descriptor_only = 1;
4821gfc_conv_expr (, arg);
4822/* This is a bare variable, so there is no
preliminary
4823   or cleanup code.  */
4824gcc_assert (se.pre.head == NULL_TREE
4825&& se.post.head == NULL_TREE);
4826rank = gfc_conv_descriptor_rank (se.expr);