[Bug lto/86321] [9 Regression] ICE: in splice_child_die, at dwarf2out.c:5693 for several Fortran tests with -g -flto

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86321

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
So in the end the issue is we have two record_types, record_type 0x7ffff6a8dc78
array01_integer(kind=4) and record_type 0x7ffff6a8de70 array01_a which
are their own TYPE_MAIN_VARIANT but share TYPE_FIELDS.

The FIELD_DECL has a DECL_CONTEXT of yet another type, record_type
0x7ffff6a8dbd0 array_descriptor01 (this is the TYPE_CANONICAL of all the
types).

This is already wrecked during LTO stream-out:

Breakpoint 5, write_ts_type_non_common_tree_pointers (ob=0x2d15ce0, 
    expr=<record_type 0x7ffff6a55498 array_descriptor01>, ref_p=false)
    at /tmp/trunk/gcc/tree-streamer-out.c:707
707         streamer_write_chain (ob, TYPE_FIELDS (expr), ref_p);
$5 = <record_type 0x7ffff6a55498 array_descriptor01>
$6 = <field_decl 0x7ffff6a531c8 data>

Breakpoint 5, write_ts_type_non_common_tree_pointers (ob=0x2d15ce0, 
    expr=<record_type 0x7ffff6a55738 array01_integer(kind=4)>, ref_p=false)
    at /tmp/trunk/gcc/tree-streamer-out.c:707
707         streamer_write_chain (ob, TYPE_FIELDS (expr), ref_p);
$7 = <record_type 0x7ffff6a55738 array01_integer(kind=4)>
$8 = <field_decl 0x7ffff6a531c8 data>

and both record types are not variants of each other.

Breakpoint 5, write_ts_type_non_common_tree_pointers (ob=0x2d15ce0, 
    expr=<record_type 0x7ffff6a591f8 array01_a>, ref_p=false)
    at /tmp/trunk/gcc/tree-streamer-out.c:707
707         streamer_write_chain (ob, TYPE_FIELDS (expr), ref_p);
$14 = <record_type 0x7ffff6a591f8 array01_a>
$15 = <field_decl 0x7ffff6a531c8 data>

It looks like this happens when doing build_distinct_type_copy () which
simply copies TYPE_FIELDs.  But I don't see how dwarf2out ever deals with
this correctly...  there's some weird unused push/pop_decl_scope () stuff
around the call to gen_member_die (scope_die_for appearantly used it at
some point).

Without LTO these array descriptor types built by
build_distinct_type_copy () never go the usual gen_member_die way
but instead they go

  /* If this is an array type with hidden descriptor, handle it first.  */
  if (!TREE_ASM_WRITTEN (type)
      && lang_hooks.types.get_array_descr_info)
    {
      memset (&info, 0, sizeof (info));
      if (lang_hooks.types.get_array_descr_info (type, &info))
        {
          /* Fortran sometimes emits array types with no dimension.  */
          gcc_assert (info.ndimensions >= 0
                      && (info.ndimensions
                          <= DWARF2OUT_ARRAY_DESCR_INFO_MAX_DIMEN));
          gen_descr_array_type_die (type, &info, context_die);
          TREE_ASM_WRITTEN (type) = 1;
          return;

which of course doesn't work when the FE is LTO.