[Bug fortran/55978] class_optional_2.f90 -Os fails

["anlauf at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55978

--- Comment #33 from anlauf at gcc dot gnu.org ---
I've been repeatedly struggling with the testcase in comment#19.

Since the dump-tree did not reveal anything, I ran a reduced version
under gdb to see why the code crashes at -O0 and -Og but not at higher
optimization.  I got them impression that the dump-tree does not tell
the full truth: subroutine one has

  {
    struct array01_integer(kind=4) * D.4292;

    D.4292 = x != 0B ? (struct array01_integer(kind=4) *)
_gfortran_internal_pack (x) : 0B;
    two (D.4292);
    if (x != 0B && (integer(kind=4)[0:] *) x->data != (integer(kind=4)[0:] *)
D.4292)
      {
        _gfortran_internal_unpack (x, D.4292);
        __builtin_free ((void *) D.4292);
      }
  }

but to my untrained eye it looked like the test x->data != D.4292 was
executed before x != 0B.  This lead me to the following tentative patch:

diff --git a/gcc/fortran/trans-array.cc b/gcc/fortran/trans-array.cc
index 19d69aec9c0..06bfea1f461 100644
--- a/gcc/fortran/trans-array.cc
+++ b/gcc/fortran/trans-array.cc
@@ -9025,7 +9100,7 @@ gfc_conv_array_parameter (gfc_se * se, gfc_expr * expr,
bool g77,
                             fold_convert (TREE_TYPE (tmp), ptr), tmp);

       if (fsym && fsym->attr.optional && sym && sym->attr.optional)
-       tmp = fold_build2_loc (input_location, TRUTH_AND_EXPR,
+       tmp = fold_build2_loc (input_location, TRUTH_ANDIF_EXPR,
                               logical_type_node,
                               gfc_conv_expr_present (sym), tmp);

Strangely enough, this gives exactly the same dump-tree!
(If someone knows why, please enlighten me!)
But it fixed the issue.

The I looked at the optimized tree.  The reference version has:

  <bb 4> [local count: 1073741824]:
  # _5 = PHI <_19(3), y_17(D)(2)>
  two (_5);
  _1 = y_17(D) != 0B;
  _2 = y_17(D)->data;
  _3 = _2 != _5;
  _4 = _1 & _3;
  if (_4 != 0)
    goto <bb 5>; [33.00%]
  else
    goto <bb 6>; [67.00%]

  <bb 5> [local count: 354334800]:
  __builtin_free (_5);

This explains the "boom".

The version with the ANDIF has:

  <bb 4> [local count: 1073741824]:
  # _2 = PHI <_16(3), y_14(D)(2)>
  two (_2);
  if (y_14(D) != 0B)
    goto <bb 5>; [70.00%]
  else
    goto <bb 7>; [30.00%]

  <bb 5> [local count: 751619280]:
  _1 = y_14(D)->data;
  if (_1 != _2)
    goto <bb 6>; [53.47%]
  else
    goto <bb 7>; [46.53%]

  <bb 6> [local count: 401890828]:
  __builtin_free (_2);

This looks sane to me: the data component is not referenced if the dummy
is seen to be absent.

Regtesting...