[Bug fortran/103691] [12 Regression] ICE in get_array_ctor_element_at_index, at fold-const.c:13314 since r12-4694-gcb153222404e2e14

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103691

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #4)
> To me it looks like the PR52329 change wasn't correct.  In particular, it
> should have been using true as the second argument and not false and
> therefore should have been removed in the r12-21-g0bf8cd9d5e8ac changes
> rather than kept.
> If I modify the testcase from a(0) to a(2), then I see
> pr103691.f90.037t.fre1:  # DEBUG D.4293 => &a[0]
> and
> pr103691.f90.038t.evrp:  # DEBUG D.4293 => &2.0e+0
> The &2.0e+0 is just a wrong-debug, debug info was supposed to contain
> address of
> a, not address of some constant that happens to be in the first element of
> the array.
> fold_stmt_1 earlier has:
>     case GIMPLE_DEBUG:
>       if (gimple_debug_bind_p (stmt))
>         {
>           tree *val = gimple_debug_bind_get_value_ptr (stmt);
>           if (*val
>               && (REFERENCE_CLASS_P (*val)
>                   || TREE_CODE (*val) == ADDR_EXPR)
>               && maybe_canonicalize_mem_ref_addr (val, true))
>             changed = true;
>         }
> which I believe should perform whatever PR52329 was meant to deal with.
> So I think
>           else if (val
>                    && TREE_CODE (val) == ADDR_EXPR)
>             {
>               tree ref = TREE_OPERAND (val, 0);
>               tree tem = maybe_fold_reference (ref);
>               if (tem)
>                 {
>                   tem = build_fold_addr_expr_with_type (tem, TREE_TYPE
> (val));
>                   gimple_debug_bind_set_value (stmt, tem);
>                   changed = true;
>                 }
>             }
> should be just dropped.

Agreed.